@misc{abadi2016deeplearningdifferentialprivacy, archiveprefix = {arXiv}, author = {Mart{\'i}n Abadi and Andy Chu and Ian Goodfellow and H. Brendan McMahan and Ilya Mironov and Kunal Talwar and Li Zhang}, doi = {10.1145/2976749.2978318}, eprint = {1607.00133}, howpublished = {arXiv preprint arXiv:1607.00133}, primaryclass = {stat.ML}, title = {{Deep Learning with Differential Privacy}}, url = {https://arxiv.org/abs/1607.00133}, year = {2016} } @inproceedings{africa2026consistency, author = {Africa, David Demitri and Mani, Arathi}, booktitle = {Proceedings of the 43rd International Conference on Machine Learning}, pages = {715--753}, publisher = {PMLR}, series = {Proceedings of Machine Learning Research}, title = {{Consistency Training Can Entrench Misalignment}}, url = {https://proceedings.mlr.press/v306/africa26a.html}, volume = {306}, year = {2026} } @inproceedings{alemohammad2024mad, archiveprefix = {arXiv}, author = {Alemohammad, Sina and Casco-Rodriguez, Josue and Luzi, Lorenzo and Humayun, Ahmed Imtiaz and Babaei, Hossein and LeJeune, Daniel and Siahkoohi, Ali and Baraniuk, Richard G.}, booktitle = {International Conference on Learning Representations}, eprint = {2307.01850}, title = {Self-Consuming Generative Models Go {MAD}}, url = {https://proceedings.iclr.cc/paper_files/paper/2024/hash/ebc042e767de551803ccfcc45e2454f5-Abstract-Conference.html}, year = {2024} } @misc{andrychowicz2016learning, archiveprefix = {arXiv}, author = {Marcin Andrychowicz and Misha Denil and Sergio Gomez and Matthew W. Hoffman and David Pfau and Tom Schaul and Brendan Shillingford and Nando de Freitas}, eprint = {1606.04474}, title = {{Learning to learn by gradient descent by gradient descent}}, url = {https://arxiv.org/abs/1606.04474}, year = {2016} } @misc{anthropic2026claudeopus55, author = {{Anthropic}}, howpublished = {Anthropic}, month = {September}, title = {{Claude Opus 5.5}}, url = {https://www.anthropic.com/claude-opus-5-5}, year = {2026} } @inproceedings{arcuschin2026wildfaithfulness, author = {Arcuschin, Iv\'{a}n and Janiak, Jett and Krzyzanowski, Robert and Rajamanoharan, Senthooran and Nanda, Neel and Conmy, Arthur}, booktitle = {Proceedings of the 43rd International Conference on Machine Learning}, pages = {3354--3405}, publisher = {PMLR}, title = {{Chain-of-Thought Reasoning In The Wild Is Not Always Faithful}}, url = {https://proceedings.mlr.press/v306/arcuschin26b.html}, volume = {306}, year = {2026} } @inproceedings{ashkinaze2025deepvalue, author = {Joshua Ashkinaze and Hua Shen and Saipranav Avula and Eric Gilbert and Ceren Budak}, booktitle = {Advances in Neural Information Processing Systems}, title = {{Deep Value Benchmark: Measuring Whether Models Generalize Deep Values or Shallow Preferences}}, url = {https://proceedings.neurips.cc/paper_files/paper/2025/hash/0f8be7310d7ed5fca40f3b133d2394e9-Abstract-Conference.html}, volume = {38}, year = {2025} } @misc{bai2022constitutional, archiveprefix = {arXiv}, author = {Yuntao Bai and Saurav Kadavath and Sandipan Kundu and Amanda Askell and Jackson Kernion and Andy Jones and Anna Chen and Anna Goldie and Azalia Mirhoseini and Cameron McKinnon and Carol Chen and Catherine Olsson and Christopher Olah and Danny Hernandez and Dawn Drain and Deep Ganguli and Dustin Li and Eli Tran-Johnson and Ethan Perez and Jamie Kerr and Jared Mueller and Jeffrey Ladish and Joshua Landau and Kamal Ndousse and Kamile Lukosuite and Liane Lovitt and Michael Sellitto and Nelson Elhage and Nicholas Schiefer and Noemi Mercado and Nova DasSarma and Robert Lasenby and Robin Larson and Sam Ringer and Scott Johnston and Shauna Kravec and Sheer El Showk and Stanislav Fort and Tamera Lanham and Timothy Telleen-Lawton and Tom Conerly and Tom Henighan and Tristan Hume and Samuel R. Bowman and Zac Hatfield-Dodds and Ben Mann and Dario Amodei and Nicholas Joseph and Sam McCandlish and Tom Brown and Jared Kaplan}, eprint = {2212.08073}, title = {{Constitutional AI: Harmlessness from AI Feedback}}, url = {https://arxiv.org/abs/2212.08073}, year = {2022} } @misc{bai2026skillzip, archiveprefix = {arXiv}, author = {Xiaofan Bai and Hongqiang Lin and Chao Liu and Yantao Zhang and Xuan Jin and Xipeng Cao and Yuhong Li}, eprint = {2608.11079}, note = {Preprint}, title = {{SkillZip: Evaluation-Free Skill Compression for Self-Evolving Agents by Discovering Reusable Structure}}, url = {https://arxiv.org/abs/2608.11079}, year = {2026} } @inproceedings{bailey2026latentobfuscation, author = {Bailey, Luke and Serrano, Alex and Sheshadri, Abhay and Seleznyov, Mikhail and Taylor, Jordan and Jenner, Erik and Hilton, Jacob and Casper, Stephen and Guestrin, Carlos and Emmons, Scott}, booktitle = {International Conference on Learning Representations}, pages = {146838--146882}, title = {{Obfuscated Activations Bypass LLM Latent-Space Defenses}}, url = {https://proceedings.iclr.cc/paper_files/paper/2026/hash/ed687a5f52b651b19e7c18f702907b8b-Abstract-Conference.html}, volume = {2026}, year = {2026} } @misc{baker2025monitoring, archiveprefix = {arXiv}, author = {Baker, Bowen and Huizinga, Joost and Gao, Leo and Dou, Zehao and Guan, Melody Y. and Madry, Aleksander and Zaremba, Wojciech and Pachocki, Jakub and Farhi, David}, eprint = {2503.11926}, title = {{Monitoring Reasoning Models for Misbehavior and the Risks of Promoting Obfuscation}}, url = {https://arxiv.org/abs/2503.11926}, year = {2025} } @misc{banerjee2026severa, archiveprefix = {arXiv}, author = {Debangshu Banerjee and Changming Xu and Eugene Ie and Ming Zhang and Daiyi Peng and Chu-Cheng Lin and Tarun Suresh and Gagandeep Singh}, eprint = {2603.25111}, howpublished = {arXiv preprint arXiv:2603.25111}, title = {{SEVerA: Verified Synthesis of Self-Evolving Agents}}, url = {https://arxiv.org/abs/2603.25111}, year = {2026} } @misc{barke2026agentrx, archiveprefix = {arXiv}, author = {Barke, Shraddha and Goyal, Arnav and Khare, Alind and Singh, Avaljot and Nath, Suman and Bansal, Chetan}, eprint = {2602.02475}, title = {{AgentRx: Diagnosing AI Agent Failures from Execution Trajectories}}, url = {https://arxiv.org/abs/2602.02475}, year = {2026} } @misc{betley2025emergent, archiveprefix = {arXiv}, author = {Jan Betley and Daniel Tan and Niels Warncke and Anna Sztyber-Betley and Xuchan Bao and Mart{\'i}n Soto and Nathan Labenz and Owain Evans}, eprint = {2502.17424}, howpublished = {arXiv preprint arXiv:2502.17424}, title = {{Emergent Misalignment: Narrow finetuning can produce broadly misaligned LLMs}}, url = {https://arxiv.org/abs/2502.17424}, year = {2025} } @inproceedings{betley2025selfreports, author = {Betley, Jan and Bao, Xuchan and Soto, Mart\'{\i}n and Sztyber-Betley, Anna and Chua, James and Evans, Owain}, booktitle = {International Conference on Learning Representations}, pages = {21127--21179}, title = {{Tell me about yourself: LLMs are aware of their learned behaviors}}, url = {https://proceedings.iclr.cc/paper_files/paper/2025/hash/364b1fd43002260dd1a9502d51d5375e-Abstract-Conference.html}, volume = {2025}, year = {2025} } @inproceedings{binder2025introspection, author = {Binder, Felix Jedidja and Chua, James and Korbak, Tomek and Sleight, Henry and Hughes, John and Long, Robert and Perez, Ethan and Turpin, Miles and Evans, Owain}, booktitle = {International Conference on Learning Representations}, pages = {3710--3756}, title = {{Looking Inward: Language Models Can Learn About Themselves by Introspection}}, url = {https://proceedings.iclr.cc/paper_files/paper/2025/hash/0a6059857ae5c82ea9726ee9282a7145-Abstract-Conference.html}, volume = {2025}, year = {2025} } @inproceedings{browncohen2024debate, author = {Brown-Cohen, Jonah and Irving, Geoffrey and Piliouras, Georgios}, booktitle = {Proceedings of the 41st International Conference on Machine Learning}, pages = {4585--4602}, publisher = {PMLR}, title = {{Scalable {AI} Safety via Doubly-Efficient Debate}}, url = {https://proceedings.mlr.press/v235/brown-cohen24a.html}, volume = {235}, year = {2024} } @inproceedings{browncohen2026proverestimator, archiveprefix = {arXiv}, author = {Jonah Brown-Cohen and Geoffrey Irving and Georgios Piliouras and Lijie Chen and Jiawei Li and Zhiyang Xun}, booktitle = {Advances in Neural Information Processing Systems}, eprint = {2506.13609}, note = {Accepted at NeurIPS 2026; proceedings forthcoming}, title = {{Avoiding Obfuscation with Prover-Estimator Debate}}, url = {https://neurips.cc/virtual/2026/poster/148404}, year = {2026} } @inproceedings{burns2024weaktostrong, author = {Burns, Collin and Izmailov, Pavel and Kirchner, Jan Hendrik and Baker, Bowen and Gao, Leo and Aschenbrenner, Leopold and Chen, Yining and Ecoffet, Adrien and Joglekar, Manas and Leike, Jan and Sutskever, Ilya and Wu, Jeffrey}, booktitle = {Proceedings of the 41st International Conference on Machine Learning}, pages = {4971--5012}, publisher = {PMLR}, title = {{Weak-to-Strong Generalization: Eliciting Strong Capabilities With Weak Supervision}}, url = {https://proceedings.mlr.press/v235/burns24b.html}, volume = {235}, year = {2024} } @misc{carlini2021extractingtrainingdatalarge, archiveprefix = {arXiv}, author = {Nicholas Carlini and Florian Tramer and Eric Wallace and Matthew Jagielski and Ariel Herbert-Voss and Katherine Lee and Adam Roberts and Tom Brown and Dawn Song and Ulfar Erlingsson and Alina Oprea and Colin Raffel}, eprint = {2012.07805}, howpublished = {arXiv preprint arXiv:2012.07805}, primaryclass = {cs.CR}, title = {{Extracting Training Data from Large Language Models}}, url = {https://arxiv.org/abs/2012.07805}, year = {2021} } @inproceedings{cemri2025mast, author = {Cemri, Mert and Pan, Melissa Z and Yang, Shuyi and Agrawal, Lakshya A and Chopra, Bhavya and Tiwari, Rishabh and Keutzer, Kurt and Parameswaran, Aditya and Klein, Dan and Ramchandran, Kannan and Zaharia, Matei A and Gonzalez, Joseph and Stoica, Ion}, booktitle = {Advances in Neural Information Processing Systems}, doi = {10.52202/085713-4082}, pages = {}, publisher = {Curran Associates, Inc.}, title = {{Why Do Multi-Agent LLM Systems Fail?}}, url = {https://proceedings.neurips.cc/paper_files/paper/2025/hash/b1041e52d3be19f0a9bc491657488e4a-Abstract-Datasets_and_Benchmarks_Track.html}, volume = {38, Main Conference}, year = {2025} } @inproceedings{chae2026modelcards, author = {Chae, Sungwon and Kim, Keonwoo and Kim, Hoki and Ju, Jaeyeon and Park, Sangchul}, booktitle = {Proceedings of the 43rd International Conference on Machine Learning}, pages = {169685--169708}, publisher = {PMLR}, title = {{Position: Current Model Cards Are Insufficient for Downstream Governance of Open-Weight Foundation Models}}, url = {https://proceedings.mlr.press/v306/chae26b.html}, volume = {306}, year = {2026} } @inproceedings{charakorn2025texttolora, author = {Rujikorn Charakorn and Edoardo Cetin and Yujin Tang and Robert Tjarko Lange}, booktitle = {Proceedings of the 42nd International Conference on Machine Learning}, pages = {7485--7514}, publisher = {PMLR}, series = {Proceedings of Machine Learning Research}, title = {{Text-to-LoRA: Instant Transformer Adaption}}, url = {https://proceedings.mlr.press/v267/charakorn25a.html}, volume = {267}, year = {2025} } @inproceedings{charakorn2026doctolora, author = {Rujikorn Charakorn and Edoardo Cetin and Shinnosuke Uesaka and Robert Tjarko Lange}, booktitle = {Proceedings of the 43rd International Conference on Machine Learning}, pages = {12967--12997}, publisher = {PMLR}, series = {Proceedings of Machine Learning Research}, title = {{Doc-to-LoRA: Learning to Instantly Internalize Contexts}}, url = {https://proceedings.mlr.press/v306/charakorn26a.html}, volume = {306}, year = {2026} } @inproceedings{chen2023controllable, author = {Chen, Sirui and Wang, Yuan and Wen, Zijing and Li, Zhiyu and Zhang, Changshuo and Zhang, Xiao and Lin, Quan and Zhu, Cheng and Xu, Jun}, booktitle = {Proceedings of the 29th ACM SIGKDD conference on knowledge discovery and data mining}, pages = {3855--3864}, title = {Controllable multi-objective re-ranking with policy hypernetworks}, year = {2023} } @inproceedings{chen2024agentpoisonredteamingllmagents, author = {Chen, Zhaorun and Xiang, Zhen and Xiao, Chaowei and Song, Dawn and Li, Bo}, booktitle = {Advances in Neural Information Processing Systems}, doi = {10.52202/079017-4136}, editor = {A. Globerson and L. Mackey and D. Belgrave and A. Fan and U. Paquet and J. Tomczak and C. Zhang}, pages = {130185--130213}, publisher = {Curran Associates, Inc.}, title = {{AgentPoison: Red-teaming LLM Agents via Poisoning Memory or Knowledge Bases}}, url = {https://proceedings.neurips.cc/paper_files/paper/2024/file/eb113910e9c3f6242541c1652e30dfd6-Paper-Conference.pdf}, volume = {37}, year = {2024} } @misc{chen2026decoevo, archiveprefix = {arXiv}, author = {Jiangwang Chen and Zixin Song and Junlin Liu and Shuaiyu Zhou and Haiyan Wu and Haihan Shi and Chenxi Zhou and Hanqing Li and Xiao Yang and Da Zhu and Guanjun Jiang and Hai Wan and Xibin Zhao}, eprint = {2607.25675}, note = {Preprint}, title = {{DecoEvo: Score-Decoupled Co-Evolution of Solver and Rubric-Generator Skills in Text Space}}, url = {https://arxiv.org/abs/2607.25675}, year = {2026} } @misc{chen2026falsefrontiers, archiveprefix = {arXiv}, author = {Meijia Chen and Hao Li and Zheng Lu and Hongshan Lin and Junbai Tian and Yichen Liu and Zijun Tian and Yufan Zou and Shuhan Sun and Hanxin Chen and Zeyu Zhang and Weizhi Du and Yueting Li and Tianyu Shi and Alaa Khamis}, eprint = {2609.39102}, title = {{False Frontiers: Diagnosing and Mitigating Co-Cheating in Self-Evolving Search Agents}}, url = {https://arxiv.org/abs/2609.39102}, year = {2026} } @misc{chen2026harnessfix, archiveprefix = {arXiv}, author = {Chen, Mengzhuo and Wang, Junjie and Liu, Zhe and Wang, Yawen and Zheng, Haiming and Wang, Qing}, eprint = {2606.06324}, title = {{From Failed Trajectories to Reliable LLM Agents: Diagnosing and Repairing Harness Flaws}}, url = {https://arxiv.org/abs/2606.06324}, year = {2026} } @inproceedings{chen2026interactiveproofs, author = {Chen, Liyan and Kalai, Yael Tauman and Xi, Zoe}, booktitle = {Proceedings of the 43rd International Conference on Machine Learning}, pages = {14831--14858}, publisher = {PMLR}, title = {{How to Avoid Debate: Scalable {AI} Safety via Doubly-Efficient Interactive Proofs}}, url = {https://proceedings.mlr.press/v306/chen26bc.html}, volume = {306}, year = {2026} } @misc{chen2026lse, archiveprefix = {arXiv}, author = {Xiaoyin Chen and Canwen Xu and Yite Wang and Boyi Liu and Zhewei Yao and Yuxiong He}, eprint = {2603.18620}, note = {Preprint}, title = {{Learning to Self-Evolve}}, url = {https://arxiv.org/abs/2603.18620}, year = {2026} } @inproceedings{chen2026mechanisticdata, author = {Chen, Jianhui and Luo, Yuzhang and Pan, Liangming}, booktitle = {Proceedings of the 43rd International Conference on Machine Learning}, pages = {15966--15998}, publisher = {PMLR}, title = {{Mechanistic Data Attribution: Tracing the Training Origins of Interpretable {LLM} Units}}, url = {https://proceedings.mlr.press/v306/chen26cv.html}, volume = {306}, year = {2026} } @inproceedings{chen2026personavectors, archiveprefix = {arXiv}, author = {Runjin Chen and Andy Arditi and Henry Sleight and Owain Evans and Jack Lindsey}, booktitle = {Advances in Neural Information Processing Systems}, eprint = {2507.21509}, note = {Accepted at NeurIPS 2026; proceedings forthcoming}, title = {{Persona Vectors: Monitoring and Controlling Character Traits in Language Models}}, url = {https://neurips.cc/virtual/2026/poster/152713}, year = {2026} } @misc{chen2026recursive, archiveprefix = {arXiv}, author = {Chen, Mingguang and Wang, Licheng and Qu, Bo}, eprint = {2607.07663}, primaryclass = {cs.AI}, title = {Recursive Self-Improvement in AI: From Bounded Self-Refinement to Autonomous Research Loops}, url = {https://arxiv.org/abs/2607.07663}, year = {2026} } @misc{chen2026safetyselfevolvingagentssurvey, archiveprefix = {arXiv}, author = {Jiahao Chen and Zhou Feng and Oubo Ma and Yichen Yan and Ruixiao Lin and Hangtao Zhang and Linkang Du and Hengyu An and Yong Yang and Jun Liu and Junhao Li and Naen Xu and Chunyi Zhou and Yuan Su and Zehao Jin and Qianli Ma and Leyi Qi and Yiming Wang and Zhe Ma and Yuwen Pu and Mengyao Du and Yuanyi Song and Enhao Huang and Zhihui Fu and Jun Wang and Jinfeng Li and Yuefeng Chen and Hui Xue and Yiming Li and Tianyu Du and Shouling Ji}, eprint = {2610.00093}, primaryclass = {cs.CR}, title = {Safety in Self-Evolving Agents: A Survey}, url = {https://arxiv.org/abs/2610.00093}, year = {2026} } @inproceedings{chen2026traceelephant, author = {Chen, Mengzhuo and Wang, Junjie and Mu, Fangwen and Wang, Yawen and Liu, Zhe and Feng, Huanxiang and Wang, Qing}, booktitle = {Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers)}, doi = {10.18653/v1/2026.acl-long.912}, pages = {19888--19905}, title = {{Seeing the Whole Elephant: A Benchmark for Failure Attribution in LLM-based Multi-Agent Systems}}, url = {https://aclanthology.org/2026.acl-long.912}, year = {2026} } @misc{chen2026whenexperiencebecomesins, archiveprefix = {arXiv}, author = {Jialuo Chen and Lingqi Jiang and Xinhao Deng and Xiaohu Du and Jianan Ma and Yunhao Feng and Yuqi Qing and Zhihao Yuan and Linkang Du and Jingyi Wang}, eprint = {2608.05563}, title = {{When Experience Becomes Instruction: Trajectory Poisoning in Self-Evolving Agent Skill Systems}}, url = {https://arxiv.org/abs/2608.05563}, year = {2026} } @misc{cheng2026contextualdrag, archiveprefix = {arXiv}, author = {Yun Cheng and Xingyu Zhu and Haoyu Zhao and Sanjeev Arora}, eprint = {2602.04288}, note = {Preprint}, title = {{Contextual Drag: How Errors in the Context Affect LLM Reasoning}}, url = {https://arxiv.org/abs/2602.04288}, year = {2026} } @inproceedings{cheng2026override, archiveprefix = {arXiv}, author = {Shuaizhi Cheng and Xiang Shi and Zhiwei Zhang and Mingwei Li}, booktitle = {Advances in Neural Information Processing Systems}, eprint = {2604.23750}, note = {Accepted at NeurIPS 2026; proceedings forthcoming}, title = {{The Override Gap: A Magnitude Account of Knowledge Conflict Failure in Hypernetwork-Based Instant LLM Adaptation}}, url = {https://neurips.cc/virtual/2026/poster/149824}, year = {2026} } @misc{cheng2026safecoevocoevolvingsafetyharnesses, archiveprefix = {arXiv}, author = {Yu Cheng and Yongkang Hu and Shuaijie Ma and Zhihang Lin and Weicheng Meng and Jingyang Qiao and Jiuan Zhou and Yushuo Zhang and Yihang Chen and Weilin Luo and Kun Shao and Dong Li and Zhizhong Zhang and Yuan Xie and Zhaoxia Yin}, eprint = {2609.36580}, primaryclass = {cs.AI}, title = {SafeCoEvo: Co-Evolving Safety Harnesses and Guards for LLM Agents at Test-Time}, url = {https://arxiv.org/abs/2609.36580}, year = {2026} } @inproceedings{chiodo2025humanloop, archiveprefix = {arXiv}, author = {Chiodo, Maurice and M\"{u}ller, Dennis and Siewert, Paul and Wetherall, Jean-Luc and Yasmine, Zoya and Burden, John}, booktitle = {International Conference on Learning Representations}, eprint = {2505.10426}, pages = {89462--89505}, title = {{Formalising Human-in-the-Loop: Computational Reductions, Failure Modes, and Legal-Moral Responsibility}}, url = {https://proceedings.iclr.cc/paper_files/paper/2026/hash/909f526db5127f8bd8158af32d9e313a-Abstract-Conference.html}, volume = {2026}, year = {2026} } @misc{clymer2024safetycases, archiveprefix = {arXiv}, author = {Clymer, Joshua and Gabrieli, Nick and Krueger, David and Larsen, Thomas}, eprint = {2403.10462}, title = {{Safety Cases: How to Justify the Safety of Advanced AI Systems}}, url = {https://arxiv.org/abs/2403.10462}, year = {2024} } @inproceedings{costa2026fides, archiveprefix = {arXiv}, author = {Manuel Costa and Boris K{\"{o}}pf and Aashish Kolluri and Andrew Paverd and Mark Russinovich and Ahmed Salem and Shruti Tople and Lukas Wutschitz and Santiago Zanella-Beguelin}, booktitle = {Advances in Neural Information Processing Systems}, eprint = {2505.23643}, note = {Accepted at NeurIPS 2026; proceedings forthcoming}, title = {{Securing AI Agents with Information-Flow Control}}, url = {https://neurips.cc/virtual/2026/poster/152696}, year = {2026} } @misc{cui2026rsianything, archiveprefix = {arXiv}, author = {Taoyong Cui and Pinxian Zeng and Wanghan Xu and Hao Xiao and Shuhan Xue and Qiang Gao and Wanli Ouyang and Pheng Ann Heng and Zhenfei Yin and Yingcheng Wu and Ling Yang}, eprint = {2610.13912}, primaryclass = {cs.CL}, title = {RSI-Anything: Unlocking Recursive Self-Improvement with World Models}, url = {https://arxiv.org/abs/2610.13912}, year = {2026} } @inproceedings{cundy2025liedetectors, author = {Chris Cundy and Adam Gleave}, booktitle = {Advances in Neural Information Processing Systems}, doi = {10.52202/085713-4313}, title = {{Preference Learning with Lie Detectors can Induce Honesty or Evasion}}, url = {https://proceedings.neurips.cc/paper_files/paper/2025/hash/bc14bd40e1186727722e23616ee2aeb9-Abstract-Conference.html}, volume = {38}, year = {2025} } @misc{dai2026recursivereward, archiveprefix = {arXiv}, author = {Dai, Wei and Liu, Qianchu and Qin, Guanghui and Han, Xing and Naumann, Tristan and Liang, Paul Pu and Zhang, Sheng}, doi = {10.48550/arXiv.2610.21040}, eprint = {2610.21040}, primaryclass = {cs.AI}, title = {Recursive Reward Improvement via Adversarial Meta-Prompt Evolution}, url = {https://arxiv.org/abs/2610.21040}, year = {2026} } @misc{dai2026recursiverewardimproveme, archiveprefix = {arXiv}, author = {Wei Dai and Qianchu Liu and Guanghui Qin and Xing Han and Tristan Naumann and Paul Pu Liang and Sheng Zhang}, eprint = {2610.21040}, title = {{Recursive Reward Improvement via Adversarial Meta-Prompt Evolution}}, url = {https://arxiv.org/abs/2610.21040}, year = {2026} } @misc{das2026seabench, archiveprefix = {arXiv}, author = {Saswat Das and Parvati Viswanathan and Daniel Donnelly and Chang Huang and Sahar Abdelnabi and Ferdinando Fioretto}, eprint = {2609.35596}, title = {{SEABench: Benchmarking Endogenous Misalignment In Self-Evolving Agents}}, url = {https://arxiv.org/abs/2609.35596}, year = {2026} } @article{dash2026memorypoisoning, author = {Dash, Pritam and Ge, Tongyu and Jain, Aditi and Shah, Tanmay and Shang, Zhiwei}, journal = {arXiv preprint arXiv:2606.04329}, title = {From Untrusted Input to Trusted Memory: A Systematic Study of Memory Poisoning Attacks in LLM Agents}, url = {https://arxiv.org/abs/2606.04329}, year = {2026} } @misc{debenedetti2024agentdojo, archiveprefix = {arXiv}, author = {Debenedetti, Edoardo and Zhang, Jie and Balunovi{\'c}, Mislav and Beurer-Kellner, Luca and Fischer, Marc and Tram{\`e}r, Florian}, eprint = {2406.13352}, title = {{AgentDojo: A Dynamic Environment to Evaluate Prompt Injection Attacks and Defenses for LLM Agents}}, url = {https://arxiv.org/abs/2406.13352}, year = {2024} } @misc{debenedetti2025defeatingpromptinjectionsdesign, archiveprefix = {arXiv}, author = {Edoardo Debenedetti and Ilia Shumailov and Tianqi Fan and Jamie Hayes and Nicholas Carlini and Daniel Fabian and Christoph Kern and Chongyang Shi and Andreas Terzis and Florian Tram{\`e}r}, eprint = {2503.18813}, howpublished = {arXiv preprint arXiv:2503.18813}, primaryclass = {cs.CR}, title = {{Defeating Prompt Injections by Design}}, url = {https://arxiv.org/abs/2503.18813}, year = {2025} } @misc{dhuliawala2023cove, archiveprefix = {arXiv}, author = {Shehzaad Dhuliawala and Mojtaba Komeili and Jing Xu and Roberta Raileanu and Xian Li and Asli Celikyilmaz and Jason Weston}, eprint = {2309.11495}, note = {Preprint}, title = {{Chain-of-Verification Reduces Hallucination in Large Language Models}}, url = {https://arxiv.org/abs/2309.11495}, year = {2023} } @misc{diao2024seas, archiveprefix = {arXiv}, author = {Muxi Diao and Rumei Li and Shiyang Liu and Guogang Liao and Jingang Wang and Xunliang Cai and Weiran Xu}, eprint = {2408.02632}, howpublished = {arXiv preprint arXiv:2408.02632}, title = {{SEAS: Self-Evolving Adversarial Safety Optimization for Large Language Models}}, url = {https://arxiv.org/abs/2408.02632}, year = {2024} } @misc{diao2026doctoatom, archiveprefix = {arXiv}, author = {Xingjian Diao and Wenbo Li and Yashas Malur Saidutta and Avinash Amballa and Lazar Valkov and Srinivas Chappidi}, eprint = {2606.12400}, title = {{Doc-to-Atom: Learning to Compile and Compose Memory Atoms}}, url = {https://arxiv.org/abs/2606.12400}, year = {2026} } @inproceedings{ding2025tailnarrowing, author = {Yiwen Ding and Zhiheng Xi and Wei He and Zhuoyuan Li and Yitao Zhai and Xiaowei Shi and Xunliang Cai and Tao Gui and Qi Zhang and Xuanjing Huang}, booktitle = {Proceedings of the 2025 Conference of the Nations of the Americas Chapter of the Association for Computational Linguistics: Human Language Technologies (Volume 1: Long Papers)}, doi = {10.18653/v1/2025.naacl-long.533}, pages = {10627--10646}, publisher = {Association for Computational Linguistics}, title = {{Mitigating Tail Narrowing in LLM Self-Improvement via Socratic-Guided Sampling}}, url = {https://aclanthology.org/2025.naacl-long.533/}, year = {2025} } @misc{ding2026distill, archiveprefix = {arXiv}, author = {Ding, Yi and Zhang, Ruqi}, eprint = {2608.31046}, title = {{Does On-Policy Distillation Really Distill? From Noisy Teacher to Self-Improvement}}, url = {https://arxiv.org/abs/2608.31046}, year = {2026} } @misc{ding2026evorubrics, archiveprefix = {arXiv}, author = {Hongxin Ding and Baixiang Huang and Yue Fang and Weibin Liao and Zheng Li and Jinyang Zhang and Zhijing Wu and Junfeng Zhao and Yasha Wang}, eprint = {2606.23038}, note = {Preprint}, title = {{EvoRubrics: Dynamic Rubrics as Rewards via Adversarial Co-Evolution for LLM Reinforcement Learning}}, url = {https://arxiv.org/abs/2606.23038}, year = {2026} } @misc{ding2026whenunlearningfails, archiveprefix = {arXiv}, author = {Zihao Ding and Jun Huang and Liang Dong}, eprint = {2607.28829}, title = {{When Unlearning Fails: Reliable Data Deletion under Post-Training in Agent Networks}}, url = {https://arxiv.org/abs/2607.28829}, year = {2026} } @misc{dobrin2026kernel, archiveprefix = {arXiv}, author = {Dobrin, Seth and Chmiel, {\L}ukasz}, eprint = {2606.26057}, title = {{The Unfireable Safety Kernel: Execution-Time AI Alignment for AI Agents and Other Escapable AI Systems}}, url = {https://arxiv.org/abs/2606.26057}, year = {2026} } @inproceedings{dong2025minja, author = {Shen Dong and Shaochen Xu and Pengfei He and Yige Li and Jiliang Tang and Tianming Liu and Hui Liu and Zhen J. Xiang}, booktitle = {Advances in Neural Information Processing Systems}, doi = {10.52202/085713-1554}, title = {Memory Injection Attacks on {LLM} Agents via Query-Only Interaction}, url = {https://proceedings.neurips.cc/paper_files/paper/2025/hash/42a97bbd9844d2bf68596730af80bcdf-Abstract-Conference.html}, volume = {38}, year = {2025} } @misc{dong2026gagar, archiveprefix = {arXiv}, author = {Dong, Jinhao and Zhao, Liang and Yue, Zihao and Ma, Wenhan and Zhang, Linghao and Li, Lei and Li, Shicheng and Song, Yifan and Ye, Bowen and Luo, Fuli}, eprint = {2609.32577}, title = {{Groupwise Agentic Grading and Advantage Redistribution for Code Agent RL}}, url = {https://arxiv.org/abs/2609.32577}, year = {2026} } @inproceedings{drayson2025detection, author = {George Drayson and Emine Yilmaz and Vasileios Lampos}, booktitle = {Proceedings of the 2025 Conference on Empirical Methods in Natural Language Processing}, doi = {10.18653/v1/2025.emnlp-main.1506}, pages = {29657--29673}, publisher = {Association for Computational Linguistics}, title = {{Machine-generated Text Detection Prevents Language Model Collapse}}, url = {https://aclanthology.org/2025.emnlp-main.1506/}, year = {2025} } @inproceedings{drori2026spillover, author = {drori, jacob and Marks, Luke and Woodworth, Bryce and Cloud, Alex and Turner, Alexander}, booktitle = {International Conference on Learning Representations}, pages = {50871--50895}, title = {{Output Supervision Can Obfuscate the Chain of Thought}}, url = {https://proceedings.iclr.cc/paper_files/paper/2026/hash/53813b70e97b6e5778e24fd88e1259f1-Abstract-Conference.html}, volume = {2026}, year = {2026} } @misc{du2026mlevolve, archiveprefix = {arXiv}, author = {Shangheng Du and Xiangchao Yan and Jinxin Shi and Zongsheng Cao and Shiyang Feng and Zichen Liang and Boyuan Sun and Tianshuo Peng and Yifan Zhou and Xin Li and Jie Zhou and Liang He and Bo Zhang and Lei Bai}, eprint = {2606.06473}, note = {Preprint}, title = {{MLEvolve: A Self-Evolving Framework for Automated Machine Learning Algorithm Discovery}}, url = {https://arxiv.org/abs/2606.06473}, year = {2026} } @misc{duan2026last, archiveprefix = {arXiv}, author = {Duan, Yi and Liu, Ying and Tang, Zirui and Chen, Haodong and Zhou, Jun and Liu, Yumou and Xu, Bangrui and Wu, Yukai and Chen, Sidi and Zhou, Yuhan and Wang, Haoyu and Yu, Xiaoyou and Han, Shaokun and Zhu, Xuzhou and Zhou, Le and Lu, Bolin and Zhou, Wei and Liu, Jiachen and Fang, Nuozhou and Tian, Jiaxin and Chen, Ruoyu and Li, Yuxuan and Zuo, Kai and Zhang, Kaiyan and Yang, Qianyu and Wang, Zijie and Qiu, Jiantao and He, Conghui and Li, Guoliang and Zhou, Bowen and Liu, Zhiyuan and Wen, Zhoufutu and Kang, Jihua and Zhou, Xuanhe and Wu, Fan}, eprint = {2609.11873}, primaryclass = {cs.LG}, title = {The Last AI Built by Humans: Toward Genuine Recursive Self-Improvement}, url = {https://arxiv.org/abs/2609.11873}, year = {2026} } @article{dwork2014foundations, author = {Cynthia Dwork and Aaron Roth}, doi = {10.1561/0400000042}, journal = {Foundations and Trends in Theoretical Computer Science}, number = {3--4}, title = {{The Algorithmic Foundations of Differential Privacy}}, url = {https://www.cis.upenn.edu/~aaroth/Papers/privacybook.pdf}, volume = {9}, year = {2014} } @article{dwork2015reusable, author = {Dwork, Cynthia and Feldman, Vitaly and Hardt, Moritz and Pitassi, Toniann and Reingold, Omer and Roth, Aaron}, doi = {10.1126/science.aaa9375}, journal = {Science}, number = {6248}, pages = {636--638}, title = {The Reusable Holdout: Preserving Validity in Adaptive Data Analysis}, url = {https://doi.org/10.1126/science.aaa9375}, volume = {349}, year = {2015} } @misc{eisenach2026l3m, archiveprefix = {arXiv}, author = {Carson Eisenach and Robert Joseph George and Vincent Quenneville-Belair and Kate Gwimm and Dean Foster}, eprint = {2610.19591}, title = {{L3M: A Verifiable Harness and Runtime for Self-Extending and Self-Improving Agents}}, url = {https://arxiv.org/abs/2610.19591}, year = {2026} } @inproceedings{engels2025oversightscaling, author = {Engels, Joshua and Baek, David and Kantamneni, Subhash and Tegmark, Max}, booktitle = {Advances in Neural Information Processing Systems}, doi = {10.52202/085713-3254}, pages = {97327--97366}, publisher = {Curran Associates, Inc.}, title = {{Scaling Laws For Scalable Oversight}}, url = {https://proceedings.neurips.cc/paper_files/paper/2025/hash/8c3caae2f725c8e2a55ecd600563d172-Abstract-Conference.html}, volume = {38, Main Conference}, year = {2025} } @misc{everitt2016selfmodification, archiveprefix = {arXiv}, author = {Tom Everitt and Daniel Filan and Mayank Daswani and Marcus Hutter}, eprint = {1605.03142}, howpublished = {arXiv preprint arXiv:1605.03142}, title = {{Self-Modification of Policy and Utility Function in Rational Agents}}, url = {https://arxiv.org/abs/1605.03142}, year = {2016} } @misc{everitt2021rewardtamperingproblemssolutions, archiveprefix = {arXiv}, author = {Tom Everitt and Marcus Hutter and Ramana Kumar and Victoria Krakovna}, eprint = {1908.04734}, howpublished = {arXiv preprint arXiv:1908.04734}, primaryclass = {cs.AI}, title = {{Reward Tampering Problems and Solutions in Reinforcement Learning: A Causal Influence Diagram Perspective}}, url = {https://arxiv.org/abs/1908.04734}, year = {2021} } @techreport{fallenstein2015vingean, author = {Fallenstein, Benja and Soares, Nate}, institution = {Machine Intelligence Research Institute}, title = {Vingean Reflection: Reliable Reasoning for Self-Improving Agents}, url = {https://intelligence.org/files/VingeanReflection.pdf}, year = {2015} } @misc{fan2026evolvingrl, archiveprefix = {arXiv}, author = {Zhiyuan Fan and Wenwei Jin and Feng Zhang and Bin Li and Yihong Dong and Yao Hu and Jiawei Li}, eprint = {2605.10663}, note = {Preprint}, title = {{Evolving-RL: End-to-End Optimization of Experience-Driven Self-Evolving Capability within Agents}}, url = {https://arxiv.org/abs/2605.10663}, year = {2026} } @misc{fan2026helix, archiveprefix = {arXiv}, author = {Tianyu Fan and Chao Huang}, eprint = {2608.13951}, title = {{HELIX: Model-Harness Co-evolution for Recursive Self-Improvement}}, url = {https://arxiv.org/abs/2608.13951}, year = {2026} } @misc{fang2025comprehensive, archiveprefix = {arXiv}, author = {Fang, Jinyuan and Peng, Yanwen and Zhang, Xi and Wang, Yingxu and Yi, Xinhao and Zhang, Guibin and Xu, Yi and Wu, Bin and Liu, Siwei and Li, Zihao and Ren, Zhaochun and Aletras, Nikos and Wang, Xi and Zhou, Han and Meng, Zaiqiao}, eprint = {2508.07407}, primaryclass = {cs.AI}, title = {A Comprehensive Survey of Self-Evolving AI Agents: A New Paradigm Bridging Foundation Models and Lifelong Agentic Systems}, url = {https://arxiv.org/abs/2508.07407}, year = {2025} } @misc{feng2026evolvingtogether, archiveprefix = {arXiv}, author = {Feng, Yujie and Wang, Xiaobin and Yang, Jinluan and Shi, Xiaonan and Xie, Pengjun and Yu, Philip S. and Wu, Xiaoming and Ye, Jieping}, doi = {10.48550/arXiv.2610.20845}, eprint = {2610.20845}, primaryclass = {cs.AI}, title = {{Evolving Together, Failing Apart}: The Skill-Policy Co-evolution Trap in Self-Improving Agents}, url = {https://arxiv.org/abs/2610.20845}, year = {2026} } @misc{feng2026evolvingtogetherfailinga, archiveprefix = {arXiv}, author = {Yujie Feng and Xiaobin Wang and Jinluan Yang and Xiaonan Shi and Pengjun Xie and Philip S. Yu and Xiaoming Wu and Jieping Ye}, eprint = {2610.20845}, title = {{Evolving Together, Failing Apart: The Skill-Policy Co-evolution Trap in Self-Improving Agents}}, url = {https://arxiv.org/abs/2610.20845}, year = {2026} } @misc{fernando2023promptbreeder, archiveprefix = {arXiv}, author = {Chrisantha Fernando and Dylan Banarse and Henryk Michalewski and Simon Osindero and Tim Rockt{\"a}schel}, eprint = {2309.16797}, title = {{Promptbreeder: Self-Referential Self-Improvement Via Prompt Evolution}}, url = {https://arxiv.org/abs/2309.16797}, year = {2023} } @inproceedings{ferreira2026causalexplanations, author = {Ferreira, Pedro and Aziz, Wilker and Titov, Ivan}, booktitle = {International Conference on Learning Representations}, pages = {130865--130894}, title = {{Truthful or Fabricated? Using Causal Attribution to Mitigate Reward Hacking in Explanations}}, url = {https://proceedings.iclr.cc/paper_files/paper/2026/hash/d4b892df8bec8e13925f57649fbae039-Abstract-Conference.html}, volume = {2026}, year = {2026} } @inproceedings{fu2025selfverification, author = {Shi Fu and Yingjie Wang and Yuzhu Chen and Li Shen and Dacheng Tao}, booktitle = {Advances in Neural Information Processing Systems}, doi = {10.52202/085713-1213}, title = {Self-Verification Provably Prevents Model Collapse in Recursive Synthetic Training}, url = {https://proceedings.neurips.cc/paper_files/paper/2025/hash/3380e8116452e0efbf36f35d95e88c94-Abstract-Conference.html}, volume = {38}, year = {2025} } @misc{fu2026sift, archiveprefix = {arXiv}, author = {Xinghong Fu and Aravinth Kulanthaivelu and Yutaro Yamada}, eprint = {2609.19526}, note = {Preprint}, title = {{Self Improvement via Fast Tree-search}}, url = {https://arxiv.org/abs/2609.19526}, year = {2026} } @misc{gao2022deletioninferencereconstructioncompliance, archiveprefix = {arXiv}, author = {Ji Gao and Sanjam Garg and Mohammad Mahmoody and Prashant Nalini Vasudevan}, eprint = {2202.03460}, howpublished = {arXiv preprint arXiv:2202.03460}, primaryclass = {cs.LG}, title = {{Deletion Inference, Reconstruction, and Compliance in Machine (Un)Learning}}, url = {https://arxiv.org/abs/2202.03460}, year = {2022} } @misc{gao2026mempoisonuncoveringpersistentmemory, archiveprefix = {arXiv}, author = {Jifeng Gao and Kang Xia and Yi Zhang and Xiaobin Hong and Mingkai Lin and Xingshen Wei and Wenzhong Li and Sanglu Lu}, eprint = {2607.14651}, howpublished = {arXiv preprint arXiv:2607.14651}, primaryclass = {cs.CR}, title = {{MemPoison: Uncovering Persistent Memory Threats and Structural Blind Spots in LLM Agents}}, url = {https://arxiv.org/abs/2607.14651}, year = {2026} } @article{gao2026selfevolving, author = {Gao, Huan-ang and Geng, Jiayi and Hua, Wenyue and Hu, Mengkang and Juan, Xinzhe and Liu, Hongzhang and Liu, Shilong and Qiu, Jiahao and Qi, Xuan and Ren, Qihan and Wu, Yiran and Wang, Hongru and Xiao, Han and Zhou, Yuhang and Zhang, Shaokun and Zhang, Jiayi and Xiang, Jinyu and Fang, Yixiong and Zhao, Qiwen and Liu, Dongrui and Qian, Cheng and Wang, Zhenhailong and Hu, Minda and Wang, Huazheng and Wu, Qingyun and Ji, Heng and Wang, Mengdi}, journal = {Transactions on Machine Learning Research}, title = {A Survey of Self-Evolving Agents: What, When, How, and Where to Evolve on the Path to Artificial Super Intelligence}, url = {https://mlanthology.org/tmlr/2026/gao2026tmlr-survey/}, year = {2026} } @misc{gerstgrasser2024accumulating, archiveprefix = {arXiv}, author = {Gerstgrasser, Matthias and Schaeffer, Rylan and Dey, Apratim and Rafailov, Rafael and Sleight, Henry and Hughes, John and Korbak, Tomasz and Agrawal, Rajashree and Pai, Dhruv and Gromov, Andrey and Roberts, Daniel A. and Yang, Diyi and Donoho, David L. and Koyejo, Sanmi}, eprint = {2404.01413}, title = {Is Model Collapse Inevitable? Breaking the Curse of Recursion by Accumulating Real and Synthetic Data}, url = {https://arxiv.org/abs/2404.01413}, year = {2024} } @inproceedings{goel2025thinkalike, author = {Goel, Shashwat and Str\"{u}ber, Joschka and Auzina, Ilze Amanda and Chandra, Karuna K and Kumaraguru, Ponnurangam and Kiela, Douwe and Prabhu, Ameya and Bethge, Matthias and Geiping, Jonas}, booktitle = {Proceedings of the 42nd International Conference on Machine Learning}, pages = {19621--19678}, publisher = {PMLR}, title = {{Great Models Think Alike and this Undermines {AI} Oversight}}, url = {https://proceedings.mlr.press/v267/goel25b.html}, volume = {267}, year = {2025} } @inproceedings{goemans2026nonmodelgains, author = {Goemans, Arthur and Altman, Daniel and Dreksler, Noemi and Schuett, Jonas and Gandhi, Milan and Wang, Zhengdong and Cogan, Sarah and Krier, Sebastien and Brady, Demetra and Ho, Lewis and Dafoe, Allan}, booktitle = {Proceedings of the 43rd International Conference on Machine Learning}, pages = {170345--170356}, publisher = {PMLR}, title = {{Position: Comprehensive {AI} governance requires addressing non-model capability gains}}, url = {https://proceedings.mlr.press/v306/goemans26a.html}, volume = {306}, year = {2026} } @inproceedings{goldowskydill2025deceptionprobes, author = {Goldowsky-Dill, Nicholas and Chughtai, Bilal and Heimersheim, Stefan and Hobbhahn, Marius}, booktitle = {Proceedings of the 42nd International Conference on Machine Learning}, pages = {19755--19786}, publisher = {PMLR}, title = {{Detecting Strategic Deception with Linear Probes}}, url = {https://proceedings.mlr.press/v267/goldowsky-dill25a.html}, volume = {267}, year = {2025} } @misc{gong2026evolutionarysafetyrecursiveselfimproving, archiveprefix = {arXiv}, author = {Chang Gong and Jingping Bi and Di Yao and Ran Cheng and Yaodong Yang and Kaile Wang and Xinjian Liang and Chao Xiang and Ruijie Guo}, eprint = {2609.31186}, primaryclass = {cs.AI}, title = {Evolutionary Safety of Recursive Self-Improving AI: Taxonomy, Risk Discovery, and Evaluation}, url = {https://arxiv.org/abs/2609.31186}, year = {2026} } @misc{gou2023critic, archiveprefix = {arXiv}, author = {Zhibin Gou and Zhihong Shao and Yeyun Gong and Yelong Shen and Yujiu Yang and Nan Duan and Weizhu Chen}, eprint = {2305.11738}, note = {Preprint}, title = {{CRITIC: Large Language Models Can Self-Correct with Tool-Interactive Critiquing}}, url = {https://arxiv.org/abs/2305.11738}, year = {2023} } @misc{greenblatt2024alignmentfaking, archiveprefix = {arXiv}, author = {Ryan Greenblatt and Carson Denison and Benjamin Wright and Fabien Roger and Monte MacDiarmid and Sam Marks and Johannes Treutlein and Tim Belonax and Jack Chen and David Duvenaud and Akbir Khan and Julian Michael and S{\"o}ren Mindermann and Ethan Perez and Linda Petrini and Jonathan Uesato and Jared Kaplan and Buck Shlegeris and Samuel R. Bowman and Evan Hubinger}, eprint = {2412.14093}, howpublished = {arXiv preprint arXiv:2412.14093}, title = {{Alignment faking in large language models}}, url = {https://arxiv.org/abs/2412.14093}, year = {2024} } @inproceedings{greenblatt2024control, author = {Greenblatt, Ryan and Shlegeris, Buck and Sachan, Kshitij and Roger, Fabien}, booktitle = {Proceedings of the 41st International Conference on Machine Learning}, pages = {16295--16336}, publisher = {PMLR}, title = {{AI Control: Improving Safety Despite Intentional Subversion}}, url = {https://proceedings.mlr.press/v235/greenblatt24a.html}, volume = {235}, year = {2024} } @inproceedings{gross2024compactproofs, author = {Gross, Jason and Agrawal, Rajashree and Kwa, Thomas and Ong, Euan and Yip, Chun Hei and Gibson, Alex and Noubir, Soufiane and Chan, Lawrence}, booktitle = {Advances in Neural Information Processing Systems}, doi = {10.52202/079017-2523}, pages = {79453--79515}, publisher = {Curran Associates, Inc.}, title = {{Compact Proofs of Model Performance via Mechanistic Interpretability}}, url = {https://proceedings.neurips.cc/paper_files/paper/2024/hash/90e73f3cf1a6c84c723a2e8b7fb2b2c1-Abstract-Conference.html}, volume = {37}, year = {2024} } @misc{gu2026copd, archiveprefix = {arXiv}, author = {Gu, Naibin and Yang, Chenxu and Si, Qingyi and Qin, Chuanyu and Yao, Dingyu and Fu, Peng and Lin, Zheng and Wang, Weiping and Duan, Nan and Wang, Jiaqi}, eprint = {2604.27083}, title = {{Co-Evolving Policy Distillation}}, url = {https://arxiv.org/abs/2604.27083}, year = {2026} } @inproceedings{guan2025monitorability, archiveprefix = {arXiv}, author = {Guan, Melody Y. and Wang, Miles and Carroll, Micah and Dou, Zehao and Wei, Annie Y. and Williams, Marcus and Arnav, Benjamin and Huizinga, Joost and Kivlichan, Ian D and Glaese, Amelia and Pachocki, Jakub and Baker, Bowen}, booktitle = {Proceedings of the 43rd International Conference on Machine Learning}, eprint = {2512.18311}, pages = {37538--37602}, publisher = {PMLR}, title = {{Monitoring Monitorability}}, url = {https://proceedings.mlr.press/v306/guan26h.html}, volume = {306}, year = {2026} } @misc{guan2025rstarmath, archiveprefix = {arXiv}, author = {Xinyu Guan and Li Lyna Zhang and Yifei Liu and Ning Shang and Youran Sun and Yi Zhu and Fan Yang and Mao Yang}, eprint = {2501.04519}, note = {Preprint}, title = {{rStar-Math: Small LLMs Can Master Math Reasoning with Self-Evolved Deep Thinking}}, url = {https://arxiv.org/abs/2501.04519}, year = {2025} } @misc{guan2026evorubric, archiveprefix = {arXiv}, author = {Xin Guan and Xiaomeng Hu and Shen Huang and Zhenyi Wang and Bo Zhang and Zijian Li and Pengjun Xie and Bo Liu and Jiuxin Cao}, eprint = {2605.29847}, note = {Preprint}, title = {{EvoRubric: Self-Evolving Rubric-Driven RL for Open-Ended Generation}}, url = {https://arxiv.org/abs/2605.29847}, year = {2026} } @misc{guo2026selfauthored, archiveprefix = {arXiv}, author = {Guo, Diandian and Cao, Cong and Yuan, Fangfang and Wang, Yingqi and Wang, Yueshan and Wang, Dakui}, eprint = {2607.24300}, primaryclass = {cs.CL}, title = {Self-Authored Verification Is Unreliable in Heuristic Self-Improving Agents}, url = {https://arxiv.org/abs/2607.24300}, year = {2026} } @inproceedings{gupta2024interpbench, author = {Gupta, Rohan and Arcuschin, Iv\'{a}n and Kwa, Thomas and Garriga-Alonso, Adri\`{a}}, booktitle = {Advances in Neural Information Processing Systems}, doi = {10.52202/079017-2950}, pages = {92922--92951}, publisher = {Curran Associates, Inc.}, title = {{InterpBench: Semi-Synthetic Transformers for Evaluating Mechanistic Interpretability Techniques}}, url = {https://proceedings.neurips.cc/paper_files/paper/2024/hash/a8f7d43ae092d9a5295775eb17f3f4f7-Abstract-Datasets_and_Benchmarks_Track.html}, volume = {37}, year = {2024} } @inproceedings{gupta2026driftaudits, author = {Gupta, Utsav}, booktitle = {Proceedings of the 43rd International Conference on Machine Learning}, pages = {170429--170439}, publisher = {PMLR}, title = {{Position: {ICML} Should Treat Hosted {LLM} {API}s as Versioned Dependencies and Require Drift-Audit Artifacts}}, url = {https://proceedings.mlr.press/v306/gupta26f.html}, volume = {306}, year = {2026} } @inproceedings{gurarieh2026faithfulnessmetrics, archiveprefix = {arXiv}, author = {Yoav Gur-Arieh and Ana Marasovic and Mor Geva}, booktitle = {Advances in Neural Information Processing Systems}, eprint = {2605.25052}, note = {Accepted at NeurIPS 2026, Evaluations and Datasets Track; proceedings forthcoming}, title = {{Faithfulness Metrics Don't Measure Faithfulness: A Meta-Evaluation with Ground Truth}}, url = {https://neurips.cc/virtual/2026/poster/139844}, year = {2026} } @inproceedings{haghtalab2026pluralistic, author = {Haghtalab, Nika and Procaccia, Ariel D. and Shao, Han and Wang, Serena Lutong and Yang, Kunhe}, booktitle = {Proceedings of the 43rd International Conference on Machine Learning}, pages = {39356--39375}, publisher = {PMLR}, series = {Proceedings of Machine Learning Research}, title = {{Pluralistic Leaderboards}}, url = {https://proceedings.mlr.press/v306/haghtalab26a.html}, volume = {306}, year = {2026} } @inproceedings{hahm2026alignmenttampering, author = {Hahm, Dongyoon and Hadfield-Menell, Dylan and Lee, Kimin}, booktitle = {Proceedings of the 43rd International Conference on Machine Learning}, pages = {39421--39450}, publisher = {PMLR}, series = {Proceedings of Machine Learning Research}, title = {{Alignment Tampering: How Reinforcement Learning from Human Feedback Is Exploited to Optimize Misaligned Biases}}, url = {https://proceedings.mlr.press/v306/hahm26a.html}, volume = {306}, year = {2026} } @inproceedings{hall2025fairnessreward, author = {Zara Hall and Melanie Subbiah and Thomas Zollo and Kathleen McKeown and Richard S. Zemel}, booktitle = {Advances in Neural Information Processing Systems}, doi = {10.52202/085713-4867}, title = {{Guiding LLM Decision-Making with Fairness Reward Models}}, url = {https://proceedings.neurips.cc/paper_files/paper/2025/hash/d6359156e0e30b1caa116a4306b12688-Abstract-Conference.html}, volume = {38}, year = {2025} } @misc{han2025alignmenttipping, archiveprefix = {arXiv}, author = {Siwei Han and Kaiwen Xiong and Jiaqi Liu and Xinyu Ye and Yaofeng Su and Wenbo Duan and Xinyuan Liu and Cihang Xie and Mohit Bansal and Mingyu Ding and Linjun Zhang and Huaxiu Yao}, eprint = {2510.04860}, howpublished = {arXiv preprint arXiv:2510.04860}, title = {{Alignment Tipping Process: How Self-Evolution Pushes LLM Agents Off the Rails}}, url = {https://arxiv.org/abs/2510.04860}, year = {2025} } @inproceedings{hashimoto2018fairness, author = {Tatsunori Hashimoto and Megha Srivastava and Hongseok Namkoong and Percy Liang}, booktitle = {Proceedings of the 35th International Conference on Machine Learning}, pages = {1929--1938}, publisher = {PMLR}, series = {Proceedings of Machine Learning Research}, title = {{Fairness Without Demographics in Repeated Loss Minimization}}, url = {https://proceedings.mlr.press/v80/hashimoto18a.html}, volume = {80}, year = {2018} } @misc{he2025aria, archiveprefix = {arXiv}, author = {He, Yufei and Li, Ruoyu and Chen, Alex and Liu, Yue and Chen, Yulin and Sui, Yuan and Chen, Cheng and Zhu, Yi and Luo, Luca and Yang, Frank and Hooi, Bryan}, eprint = {2507.17131}, title = {{Enabling Self-Improving Agents to Learn at Test Time With Human-In-The-Loop Guidance}}, url = {https://arxiv.org/abs/2507.17131}, year = {2025} } @misc{hebbar2026sia, archiveprefix = {arXiv}, author = {Prannay Hebbar and Yogendra Manawat and Samuel Verboomen and Alesia Ivanova and Selvam Palanimalai and Kunal Bhatia and Vignesh Baskaran}, eprint = {2605.27276}, note = {Preprint}, title = {{SIA: Self Improving AI with Harness \& Weight Updates}}, url = {https://arxiv.org/abs/2605.27276}, year = {2026} } @techreport{hleg2019ethics, author = {{High-Level Expert Group on Artificial Intelligence}}, institution = {European Commission}, month = {April}, title = {Ethics Guidelines for Trustworthy {AI}}, url = {https://digital-strategy.ec.europa.eu/en/library/ethics-guidelines-trustworthy-ai}, year = {2019} } @misc{hosseini2024vstar, archiveprefix = {arXiv}, author = {Arian Hosseini and Xingdi Yuan and Nikolay Malkin and Aaron Courville and Alessandro Sordoni and Rishabh Agarwal}, eprint = {2402.06457}, note = {Preprint}, title = {{V-STaR: Training Verifiers for Self-Taught Reasoners}}, url = {https://arxiv.org/abs/2402.06457}, year = {2024} } @inproceedings{hosseini2025distributive, author = {Hadi Hosseini and Samarth Khanna}, booktitle = {Advances in Neural Information Processing Systems}, title = {{Distributive Fairness in Large Language Models: Evaluating Alignment with Human Values}}, url = {https://proceedings.neurips.cc/paper_files/paper/2025/hash/7d411ed79f762a0e99337cf4ce38938a-Abstract-Conference.html}, volume = {38}, year = {2025} } @inproceedings{hsu2024safelora, author = {Chia-Yi Hsu and Yu-Lin Tsai and Chih-Hsun Lin and Pin-Yu Chen and Chia-Mu Yu and Chun-Ying Huang}, booktitle = {Advances in Neural Information Processing Systems}, title = {{Safe LoRA: the Silver Lining of Reducing Safety Risks when Fine-tuning Large Language Models}}, url = {https://arxiv.org/abs/2405.16833}, volume = {37}, year = {2024} } @inproceedings{hua2025costmonitors, author = {Hua, Tim and Baskerville, James and Lemoine, Henri and Hopman, Mia and Bhatt, Aryan and Tracy, Tyler}, booktitle = {Advances in Neural Information Processing Systems}, doi = {10.52202/085713-5416}, pages = {162175--162200}, publisher = {Curran Associates, Inc.}, title = {{Combining Cost Constrained Runtime Monitors for AI Safety}}, url = {https://proceedings.neurips.cc/paper_files/paper/2025/hash/ed56a654ddb681841fdb2184463c253e-Abstract-Conference.html}, volume = {38, Main Conference}, year = {2025} } @inproceedings{hua2026evaluationaware, author = {Hua, Tim and Qin, Andrew and Marks, Samuel and Nanda, Neel}, booktitle = {International Conference on Learning Representations}, pages = {90908--90968}, title = {{Steering Evaluation-Aware Language Models To Act Like They Are Deployed}}, url = {https://proceedings.iclr.cc/paper_files/paper/2026/hash/9334fd3a5170dbfe74eae4755f6c5f89-Abstract-Conference.html}, volume = {2026}, year = {2026} } @inproceedings{huang2024selfcorrect, archiveprefix = {arXiv}, author = {Huang, Jie and Chen, Xinyun and Mishra, Swaroop and Zheng, Huaixiu Steven and Yu, Adams Wei and Song, Xinying and Zhou, Denny}, booktitle = {International Conference on Learning Representations}, eprint = {2310.01798}, title = {Large Language Models Cannot Self-Correct Reasoning Yet}, url = {https://proceedings.iclr.cc/paper_files/paper/2024/hash/8b4add8b0aa8749d80a34ca5d941c355-Abstract-Conference.html}, year = {2024} } @misc{huang2024sharpening, archiveprefix = {arXiv}, author = {Audrey Huang and Adam Block and Dylan J. Foster and Dhruv Rohatgi and Cyril Zhang and Max Simchowitz and Jordan T. Ash and Akshay Krishnamurthy}, eprint = {2412.01951}, note = {Preprint}, title = {{Self-Improvement in Language Models: The Sharpening Mechanism}}, url = {https://arxiv.org/abs/2412.01951}, year = {2024} } @inproceedings{huang2024trustllm, author = {Huang, Yue and Sun, Lichao and Wang, Haoran and Wu, Siyuan and Zhang, Qihui and Li, Yuan and Gao, Chujie and Huang, Yixin and Lyu, Wenhan and Zhang, Yixuan and Li, Xiner and Sun, Hanchi and Liu, Zhengliang and Liu, Yixin and Wang, Yijue and Zhang, Zhikun and Vidgen, Bertie and Kailkhura, Bhavya and Xiong, Caiming and Xiao, Chaowei and Li, Chunyuan and Xing, Eric P. and Huang, Furong and Liu, Hao and Ji, Heng and Wang, Hongyi and Zhang, Huan and Yao, Huaxiu and Kellis, Manolis and Zitnik, Marinka and Jiang, Meng and Bansal, Mohit and Zou, James and Pei, Jian and Liu, Jian and Gao, Jianfeng and Han, Jiawei and Zhao, Jieyu and Tang, Jiliang and Wang, Jindong and Vanschoren, Joaquin and Mitchell, John and Shu, Kai and Xu, Kaidi and Chang, Kai-Wei and He, Lifang and Huang, Lifu and Backes, Michael and Gong, Neil Zhenqiang and Yu, Philip S. and Chen, Pin-Yu and Gu, Quanquan and Xu, Ran and Ying, Rex and Ji, Shuiwang and Jana, Suman and Chen, Tianlong and Liu, Tianming and Zhou, Tianyi and Wang, William Yang and Li, Xiang and Zhang, Xiangliang and Wang, Xiao and Xie, Xing and Chen, Xun and Wang, Xuyu and Liu, Yan and Ye, Yanfang and Cao, Yinzhi and Chen, Yong and Zhao, Yue}, booktitle = {Proceedings of the 41st International Conference on Machine Learning}, pages = {20166--20270}, publisher = {PMLR}, series = {Proceedings of Machine Learning Research}, title = {Position: {TrustLLM}: Trustworthiness in Large Language Models}, url = {https://proceedings.mlr.press/v235/huang24x.html}, volume = {235}, year = {2024} } @misc{huang2026reward, archiveprefix = {arXiv}, author = {Huang, Yue and Xu, Zhangchen and Ma, Yuchen and Wang, Wenjie and Liu, Zheyuan and Xu, Ziwei and Chen, Pin-Yu and Galley, Michel and Lin, Zinan and Feuerriegel, Stefan and Poovendran, Radha and Sra, Misha and Pentland, Alex and Zhang, Xiangliang and Chen, Zichen}, eprint = {2609.28614}, note = {Preprint}, title = {{Reward Hacking Challenges Oversight of Autonomous Research Agents}}, url = {https://arxiv.org/abs/2609.28614}, year = {2026} } @inproceedings{hudson2026corrigibility, author = {Hudson, Rubi}, booktitle = {Proceedings of the 43rd International Conference on Machine Learning}, pages = {48722--48745}, publisher = {PMLR}, title = {{Corrigibility Transformation: Constructing Goals That Accept Updates}}, url = {https://proceedings.mlr.press/v306/hudson26a.html}, volume = {306}, year = {2026} } @misc{iacob2026redqueen, archiveprefix = {arXiv}, author = {Alex Iacob and Andrej Jovanović and William F. Shen and Daniel Burkhardt and Meghdad Kurmanji and Nurbek Tastan and Lorenzo Sani and Niccolò Alberto Elia Venanzi and Jiayi Nie and Ambroise Odonnat and Zeyu Cao and Bill Marino and Xinchi Qiu and Rika Antonova and Nicholas D. Lane}, eprint = {2606.26294}, primaryclass = {cs.LG}, title = {The Red Queen G\"odel Machine: Co-Evolving Agents and Their Evaluators}, url = {https://arxiv.org/abs/2606.26294}, year = {2026} } @inproceedings{jeon2025w2sshift, author = {Jeon, Myeongho and Sobotka, Jan and Choi, Suhwan and Brbic, Maria}, booktitle = {Advances in Neural Information Processing Systems}, doi = {10.52202/085713-5130}, pages = {153420--153458}, publisher = {Curran Associates, Inc.}, title = {{Weak-to-Strong Generalization under Distribution Shifts}}, url = {https://proceedings.neurips.cc/paper_files/paper/2025/hash/e19583fd7768217b49db8f77c9f3e952-Abstract-Conference.html}, volume = {38, Main Conference}, year = {2025} } @misc{jiang2025aide, archiveprefix = {arXiv}, author = {Zhengyao Jiang and Dominik Schmidt and Dhruv Srikanth and Dixing Xu and Ian Kaplan and Deniss Jacenko and Yuxiang Wu}, eprint = {2502.13138}, note = {Preprint}, title = {{AIDE: AI-Driven Exploration in the Space of Code}}, url = {https://arxiv.org/abs/2502.13138}, year = {2025} } @misc{jiang2026agentlabbenchmarkingllmagents, archiveprefix = {arXiv}, author = {Tanqiu Jiang and Yuhui Wang and Jiacheng Liang and Ting Wang}, eprint = {2602.16901}, howpublished = {arXiv preprint arXiv:2602.16901}, primaryclass = {cs.AI}, title = {{AgentLAB: Benchmarking LLM Agents against Long-Horizon Attacks}}, url = {https://arxiv.org/abs/2602.16901}, year = {2026} } @inproceedings{jiang2026disagreement, author = {Jiang, Yuyang and Chen, Chacha and Wu, Teng and Sun, Liwen and Liu, Han and Feng, Shi and Tan, Chenhao}, booktitle = {Proceedings of the 43rd International Conference on Machine Learning}, pages = {52111--52137}, publisher = {PMLR}, title = {{Collaborative Disagreement Resolution for Scalable Oversight}}, url = {https://proceedings.mlr.press/v306/jiang26d.html}, volume = {306}, year = {2026} } @misc{jiang2026harnessevolve, archiveprefix = {arXiv}, author = {Wen Jiang and Mingmin Chu and Yimeng Tian and Qianxin Zhang and Haofei Yang and Rui Yang and Yang Liu and Tao Lv and Fangming Li}, eprint = {2609.00829}, note = {Preprint}, title = {{HarnessEvolve: Learning from Reference Trajectories for Reliable Agent Self-Evolution}}, url = {https://arxiv.org/abs/2609.00829}, year = {2026} } @misc{jiang2026seaeval, archiveprefix = {arXiv}, author = {Sihang Jiang and Lipeng Ma and Zhonghua Hong and Keyi Wang and Zhiyu Lu and Tengfei Wang and Shisong Chen and Jinghao Zhang and Tianjun Pan and Weijia Li and Jiaqing Liang and Yanghua Xiao}, eprint = {2604.08988}, note = {Preprint}, title = {{SEA-Eval: A Benchmark for Evaluating Self-Evolving Agents Beyond Episodic Assessment}}, url = {https://arxiv.org/abs/2604.08988}, year = {2026} } @article{jiang2026selfimproving, author = {Jiang, Che and Zhong, Jincheng and Fu, Yu and Tian, Kai and Yang, Junlin and Zhao, Kaikai and Wang, Yuchong and Luo, Tianwei and Wang, Weizhi and Zuo, Yuxin and Jia, Guoli and Lv, Xingtai and Lei, Dianqiao and Zeng, Sihang and Wang, Yuru and Yuan, Zhenzhao and Long, Xinwei and Hua, Ermo and Ren, Can and Jiang, Xin and Xie, Shulei and Zheng, Yuanchun and Sun, Youbang and Qi, Biqing and Ding, Ning and Zhang, Kaiyan and Zhou, Bowen}, journal = {OpenReview Archive}, title = {Self-Improving Agents in the Era of Experience: A Survey of Self- to Meta-Evolution}, url = {https://openreview.net/forum?id=IUltZSgLMm}, year = {2026} } @misc{jiang2026sorsi, archiveprefix = {arXiv}, author = {Jiang, Yuxuan and Vempaty, Aditya and Jagmohan, Ashish}, eprint = {2610.05701}, primaryclass = {cs.AI}, title = {Second-Order Problem Solving for Recursive Self-Improvement in Formal Verification}, url = {https://arxiv.org/abs/2610.05701}, year = {2026} } @misc{jin2025reveal, archiveprefix = {arXiv}, author = {Yiyang Jin and Kunzhao Xu and Hang Li and Xueting Han and Yanmin Zhou and Cheng Li and Jing Bai}, eprint = {2506.11442}, note = {Preprint}, title = {{ReVeal: Self-Evolving Code Agents via Reliable Self-Verification}}, url = {https://arxiv.org/abs/2506.11442}, year = {2025} } @misc{johnson2026aimprivacyawareinteroperablememory, archiveprefix = {arXiv}, author = {Zachary Johnson and Nigel Boachie Kumankumah and Somya Chatterjee and Tejas Sathyamurthi and Min Chen and Xinyi Alice Li and Xiao Wang and Emily Morgan Gelchie and Jessica Lin and Sadid A. Hasan and Sulaiman Vesal}, eprint = {2609.12320}, howpublished = {arXiv preprint arXiv:2609.12320}, primaryclass = {cs.AI}, title = {{AIM: A Privacy-Aware Interoperable Memory Framework for Multi-Agent Multi-User LLM Systems}}, url = {https://arxiv.org/abs/2609.12320}, year = {2026} } @inproceedings{jung2025escalate, author = {Jung, Jaehun and Brahman, Faeze and Choi, Yejin}, booktitle = {International Conference on Learning Representations}, pages = {3101--3125}, title = {{Trust or Escalate: LLM Judges with Provable Guarantees for Human Agreement}}, url = {https://proceedings.iclr.cc/paper_files/paper/2025/hash/08dabd5345b37fffcbe335bd578b15a0-Abstract-Conference.html}, volume = {2025}, year = {2025} } @misc{kadavath2022know, archiveprefix = {arXiv}, author = {Saurav Kadavath and Tom Conerly and Amanda Askell and Tom Henighan and Dawn Drain and Ethan Perez and Nicholas Schiefer and Zac Hatfield-Dodds and Nova DasSarma and Eli Tran-Johnson and Scott Johnston and Sheer El-Showk and Andy Jones and Nelson Elhage and Tristan Hume and Anna Chen and Yuntao Bai and Sam Bowman and Stanislav Fort and Deep Ganguli and Danny Hernandez and Josh Jacobson and Jackson Kernion and Shauna Kravec and Liane Lovitt and Kamal Ndousse and Catherine Olsson and Sam Ringer and Dario Amodei and Tom Brown and Jack Clark and Nicholas Joseph and Ben Mann and Sam McCandlish and Chris Olah and Jared Kaplan}, eprint = {2207.05221}, note = {Preprint}, title = {{Language Models (Mostly) Know What They Know}}, url = {https://arxiv.org/abs/2207.05221}, year = {2022} } @inproceedings{kale2026monitoring, archiveprefix = {arXiv}, author = {Neil Kale and Chen Bo Calvin Zhang and Kevin Zhu and Ankit Aich and Paula Rodriguez and {Scale Red Team} and Christina Q. Knight and Zifan Wang}, booktitle = {International Conference on Learning Representations}, eprint = {2508.19461}, title = {{Reliable Weak-to-Strong Monitoring of LLM Agents}}, url = {https://arxiv.org/abs/2508.19461}, year = {2026} } @misc{karten2026continualharness, archiveprefix = {arXiv}, author = {Seth Karten and Joel Zhang and Tersoo Upaa and Ruirong Feng and Wenzhe Li and Chengshuai Shi and Chi Jin and Kiran Vodrahalli}, eprint = {2605.09998}, note = {Preprint}, title = {{Continual Harness: Online Adaptation for Self-Improving Foundation Agents}}, url = {https://arxiv.org/abs/2605.09998}, year = {2026} } @inproceedings{karvonen2025saebench, author = {Karvonen, Adam and Rager, Can and Lin, Johnny and Tigges, Curt and Bloom, Joseph Isaac and Chanin, David and Lau, Yeu-Tong and Farrell, Eoin and Mcdougall, Callum Stuart and Ayonrinde, Kola and Till, Demian and Wearden, Matthew and Conmy, Arthur and Marks, Samuel and Nanda, Neel}, booktitle = {Proceedings of the 42nd International Conference on Machine Learning}, pages = {29223--29264}, publisher = {PMLR}, title = {{{SAEB}ench: A Comprehensive Benchmark for Sparse Autoencoders in Language Model Interpretability}}, url = {https://proceedings.mlr.press/v267/karvonen25a.html}, volume = {267}, year = {2025} } @inproceedings{kenton2024weakjudges, author = {Kenton, Zachary and Siegel, Noah Y. and Kram\'{a}r, J\'{a}nos and Brown-Cohen, Jonah and Albanie, Samuel and Bulian, Jannis and Agarwal, Rishabh and Lindner, David and Tang, Yunhao and Goodman, Noah and Shah, Rohin}, booktitle = {Advances in Neural Information Processing Systems}, doi = {10.52202/079017-2395}, pages = {75229--75276}, publisher = {Curran Associates, Inc.}, title = {{On scalable oversight with weak LLMs judging strong LLMs}}, url = {https://proceedings.neurips.cc/paper_files/paper/2024/hash/899511e37a8e01e1bd6f6f1d377cc250-Abstract-Conference.html}, volume = {37}, year = {2024} } @inproceedings{khalaf2025inferencerewardhacking, author = {Hadi Khalaf and Claudio Mayrink Verdun and Alex Oesterling and Himabindu Lakkaraju and Flavio Calmon}, booktitle = {Advances in Neural Information Processing Systems}, doi = {10.52202/085713-2064}, title = {{Inference-Time Reward Hacking in Large Language Models}}, url = {https://proceedings.neurips.cc/paper_files/paper/2025/hash/590a0cc0306c1c63e2d66a51a407718f-Abstract-Conference.html}, volume = {38}, year = {2025} } @misc{kirk2024prism, archiveprefix = {arXiv}, author = {Hannah Rose Kirk and Alexander Whitefield and Paul Röttger and Andrew Bean and Katerina Margatina and Juan Ciro and Rafael Mosquera and Max Bartolo and Adina Williams and He He and Bertie Vidgen and Scott A. Hale}, eprint = {2404.16019}, primaryclass = {cs.CL}, title = {The PRISM Alignment Dataset: What Participatory, Representative and Individualised Human Feedback Reveals About the Subjective and Multicultural Alignment of Large Language Models}, url = {https://arxiv.org/abs/2404.16019}, year = {2024} } @inproceedings{knecht2026awarenesseffects, archiveprefix = {arXiv}, author = {Amelie Knecht and Lucas Florin and Thilo Hagendorff}, booktitle = {Advances in Neural Information Processing Systems}, eprint = {2605.05835}, note = {Accepted at NeurIPS 2026; proceedings forthcoming}, title = {{Evaluation Awareness in Language Models Has Limited Effect on Behaviour}}, url = {https://neurips.cc/virtual/2026/poster/148437}, year = {2026} } @inproceedings{kumar2025score, author = {Aviral Kumar and Vincent Zhuang and Rishabh Agarwal and Yi Su and JD Co-Reyes and Avi Singh and Kate Baumli and Shariq Iqbal and Colton Bishop and Rebecca Roelofs and Lei Zhang and Kay McKinney and Disha Shrivastava and Cosmin Paduraru and George Tucker and Doina Precup and Feryal Behbahani and Aleksandra Faust}, booktitle = {International Conference on Learning Representations}, title = {{Training Language Models to Self-Correct via Reinforcement Learning}}, url = {https://proceedings.iclr.cc/paper_files/paper/2025/hash/871ac99fdc5282d0301934d23945ebaa-Abstract-Conference.html}, year = {2025} } @inproceedings{lai2026reinforcement, author = {Lai, Song and Zhao, Haohan and Feng, Rong and Ma, Changyi and Liu, Wenzhuo and Zhao, Hongbo and Lin, Xi and Yi, Dong and Zhang, Qingfu and Liu, Hongbin and Meng, Gaofeng and Zhu, Fei}, booktitle = {Proceedings of the 43rd International Conference on Machine Learning}, pages = {62363--62378}, publisher = {PMLR}, series = {Proceedings of Machine Learning Research}, title = {Reinforcement Fine-Tuning Naturally Mitigates Forgetting in Continual Post-Training}, url = {https://proceedings.mlr.press/v306/lai26h.html}, volume = {306}, year = {2026} } @inproceedings{langosco2022goalmisgeneralization, author = {Langosco, Lauro Langosco Di and Koch, Jack and Sharkey, Lee D and Pfau, Jacob and Krueger, David}, booktitle = {Proceedings of the 39th International Conference on Machine Learning}, pages = {12004--12019}, publisher = {PMLR}, series = {Proceedings of Machine Learning Research}, title = {{Goal Misgeneralization in Deep Reinforcement Learning}}, url = {https://proceedings.mlr.press/v162/langosco22a.html}, volume = {162}, year = {2022} } @misc{lee2026metaharness, archiveprefix = {arXiv}, author = {Yoonho Lee and Roshen Nair and Qizheng Zhang and Kangwook Lee and Omar Khattab and Chelsea Finn}, eprint = {2603.28052}, title = {{Meta-Harness: End-to-End Optimization of Model Harnesses}}, url = {https://arxiv.org/abs/2603.28052}, year = {2026} } @misc{lee2026recursiveselfimprovement, archiveprefix = {arXiv}, author = {Hyunin Lee and Jinglue Xu and Jeffrey Seely and Donghyun Lee and Somayeh Sojoudi and Matei Zaharia and Yujin Tang}, eprint = {2610.12176}, title = {{Recursive Self-Improvement through Multi-Agent Self-Supervision}}, url = {https://arxiv.org/abs/2610.12176}, year = {2026} } @misc{lee2026selfevolvingsearchindex, archiveprefix = {arXiv}, author = {Sangam Lee and Wonjae Lee and Sunghwan Kim and Deogyong Kim and Jaehoon Kim and Daye Nam and SeongKu Kang and Dongha Lee}, eprint = {2609.19656}, primaryclass = {cs.IR}, title = {Self-Evolving Search Index}, url = {https://arxiv.org/abs/2609.19656}, year = {2026} } @article{li2023trustworthy, articleno = {177}, author = {Bo Li and Peng Qi and Bo Liu and Shuai Di and Jingen Liu and Jiquan Pei and Jinfeng Yi and Bowen Zhou}, doi = {10.1145/3555803}, journal = {ACM Computing Surveys}, number = {9}, numpages = {46}, title = {Trustworthy {AI}: From Principles to Practices}, url = {https://doi.org/10.1145/3555803}, volume = {55}, year = {2023} } @misc{li2026evolm, archiveprefix = {arXiv}, author = {Shuyue Stella Li and Rui Xin and Teng Xiao and Yike Wang and Rulin Shao and Zoey Hao and Melanie Sclar and Sewoong Oh and Faeze Brahman and Pang Wei Koh and Yulia Tsvetkov}, eprint = {2605.03871}, note = {Preprint}, title = {{EvoLM: Self-Evolving Language Models through Co-Evolved Discriminative Rubrics}}, url = {https://arxiv.org/abs/2605.03871}, year = {2026} } @misc{li2026memoryechoinsingleagentl, archiveprefix = {arXiv}, author = {Yufeng Li and Sheng Yao and Peng Zeng and Dezhang Kong and Shi Lin and Changting Lin and Xun Wang and Meng Han}, eprint = {2610.17763}, title = {{Memory Echo in Single-Agent Long-Term Memory: Lifecycle Evaluation and Provenance-Guided Mitigation}}, url = {https://arxiv.org/abs/2610.17763}, year = {2026} } @misc{li2026rst, archiveprefix = {arXiv}, author = {Zhongzhi Li and Yucheng Shi and Zongxia Li and Ruhan Wang and Anhao Li and Zixun Huang and Junyao Yang and Lei Ke and Ninghao Liu and Haitao Mi and Leowei Liang}, eprint = {2608.05466}, note = {Preprint}, title = {{Recursive Synthesis for Long-Horizon Terminal Tasks}}, url = {https://arxiv.org/abs/2608.05466}, year = {2026} } @misc{li2026sesearchselfevolvingsearchagent, archiveprefix = {arXiv}, author = {Jian Li and Yizhang Jin and Dongqi Liu and Hang Ding and Jiafu Wu and Dongsheng Chen and Yunhang Shen and Yulei Qin and Ying Tai and Chengjie Wang and Xiaotong Yuan and Yabiao Wang}, eprint = {2603.03293}, primaryclass = {cs.CL}, title = {SE-Search: Self-Evolving Search Agent via Memory and Dense Reward}, url = {https://arxiv.org/abs/2603.03293}, year = {2026} } @misc{li2026traceevolutionbench, archiveprefix = {arXiv}, author = {Fangfei Li and Chenyang Zhao and Long Wang and Feng Tian and Lv Guo}, eprint = {2610.15055}, title = {{Trace Evolution Bench: Measuring Whether Agents Improve from Their Own Environment Traces}}, url = {https://arxiv.org/abs/2610.15055}, year = {2026} } @misc{liang2022holistic, archiveprefix = {arXiv}, author = {Liang, Percy and Bommasani, Rishi and Lee, Tony and Tsipras, Dimitris and Soylu, Dilara and Yasunaga, Michihiro and Zhang, Yian and Narayanan, Deepak and Wu, Yuhuai and Kumar, Ananya and Newman, Benjamin and Yuan, Binhang and Yan, Bobby and Zhang, Ce and Cosgrove, Christian and Manning, Christopher D. and Ré, Christopher and Acosta-Navas, Diana and Hudson, Drew A. and Zelikman, Eric and Durmus, Esin and Ladhak, Faisal and Rong, Frieda and Ren, Hongyu and Yao, Huaxiu and Wang, Jue and Santhanam, Keshav and Orr, Laurel and Zheng, Lucia and Yuksekgonul, Mert and Suzgun, Mirac and Kim, Nathan and Guha, Neel and Chatterji, Niladri and Khattab, Omar and Henderson, Peter and Huang, Qian and Chi, Ryan and Xie, Sang Michael and Santurkar, Shibani and Ganguli, Surya and Hashimoto, Tatsunori and Icard, Thomas and Zhang, Tianyi and Chaudhary, Vishrav and Wang, William and Li, Xuechen and Mai, Yifan and Zhang, Yuhui and Koreeda, Yuta}, eprint = {2211.09110}, title = {{Holistic Evaluation of Language Models}}, url = {https://arxiv.org/abs/2211.09110}, year = {2022} } @misc{liang2026robofoundrysystemaspolicyevolutionselflearning, archiveprefix = {arXiv}, author = {Jingsong Liang and Shuhao Liao and Shizhe Zhang and Diyuan Hou and Yuxin Cai and Xinjian Deng and Chengyang He and Wenhui Huang and Runjia Tan and Zhidong Wang and Lan Yu and Xuesong Tian and Guillaume Sartoretti and Jie Luo and Yao Mu and Wenjun Wu and Wanhua Li and Chen Lv}, eprint = {2609.32862}, primaryclass = {cs.RO}, title = {RoboFoundry: System-as-Policy Evolution for Self-Learning Embodied Agents}, url = {https://arxiv.org/abs/2609.32862}, year = {2026} } @misc{lightman2023processsupervision, archiveprefix = {arXiv}, author = {Hunter Lightman and Vineet Kosaraju and Yura Burda and Harri Edwards and Bowen Baker and Teddy Lee and Jan Leike and John Schulman and Ilya Sutskever and Karl Cobbe}, eprint = {2305.20050}, note = {Preprint}, title = {{Let's Verify Step by Step}}, url = {https://arxiv.org/abs/2305.20050}, year = {2023} } @misc{lin2026ahe, archiveprefix = {arXiv}, author = {Lin, Jiahang and Liu, Shichun and Pan, Chengjun and Lin, Lizhi and Dou, Shihan and Xi, Zhiheng and Huang, Xuanjing and Yan, Hang and Han, Zhenhua and Gui, Tao and Jiang, Yu-Gang}, eprint = {2604.25850}, title = {{Agentic Harness Engineering: Observability-Driven Automatic Evolution of Coding-Agent Harnesses}}, url = {https://arxiv.org/abs/2604.25850}, year = {2026} } @misc{lin2026amnesia, archiveprefix = {arXiv}, author = {Jianzhe Lin and Fei Wang and Xiaolin Li and Rajeshkumar Golani and Jubin Chheda}, eprint = {2606.21089}, note = {Preprint}, title = {{Repeated post-training is not Self-improving: Diagnosing Scientific Amnesia in Continual DPO Pipelines}}, url = {https://arxiv.org/abs/2606.21089}, year = {2026} } @misc{lin2026beyondendpointperformanc, archiveprefix = {arXiv}, author = {Hongqiang Lin and Chao Liu and Xiaofan Bai and Xuan Jin and Yuhong Li and Nenggan Zheng and Xipeng Cao}, eprint = {2609.24663}, title = {{Beyond Endpoint Performance: Process-Level Evaluation of Self-Evolving Agents}}, url = {https://arxiv.org/abs/2609.24663}, year = {2026} } @misc{lin2026canagentstrulyforgetexpl, archiveprefix = {arXiv}, author = {Junkai Lin and Ge Xu and Zheng Xu and Junkai Chen and Xiangyuan Lan}, eprint = {2610.17659}, title = {{Can Agents Truly Forget? Explore Machine Unlearning for LLM Agents}}, url = {https://arxiv.org/abs/2610.17659}, year = {2026} } @misc{lin2026harnessbenefit, archiveprefix = {arXiv}, author = {Minhua Lin and Juncheng Wu and Zijun Wang and Zhan Shi and Yisi Sang and Bing He and Zewen Liu and Tianxin Wei and Zongyu Wu and Zhiwei Zhang and Dakuo Wang and Xiang Zhang and Benoit Dumoulin and Cihang Xie and Yuyin Zhou and Suhang Wang and Hanqing Lu}, eprint = {2605.30621}, title = {{Harness Updating Is Not Harness Benefit: Disentangling Evolution Capabilities in Self-Evolving LLM Agents}}, url = {https://arxiv.org/abs/2605.30621}, year = {2026} } @misc{lin2026safetyselfevolvingllmagent, archiveprefix = {arXiv}, author = {Ruixiao Lin and Xinhao Deng and Qingming Li and Jianan Ma and Yunhao Feng and Yuqi Qing and Zhenyuan Li and Yechao Zhang and Shiwen Cui and Changhua Meng and Tianwei Zhang and Xingjun Ma and Qi Li and Ke Xu and Shouling Ji}, eprint = {2606.23075}, howpublished = {arXiv preprint arXiv:2606.23075}, primaryclass = {cs.CR}, title = {{Safety in Self-Evolving LLM Agent Systems: Threats, Amplification, and Case Studies}}, url = {https://arxiv.org/abs/2606.23075}, year = {2026} } @misc{lin2026selfregress, archiveprefix = {arXiv}, author = {Lin, Jianzhe}, eprint = {2606.21090}, note = {Preprint}, title = {Self-Improvement Can Self-Regress: The Rise-and-Collapse Failure Mode of {LLM} Self-Training}, url = {https://arxiv.org/abs/2606.21090}, year = {2026} } @inproceedings{liu2018delayed, author = {Lydia T. Liu and Sarah Dean and Esther Rolf and Max Simchowitz and Moritz Hardt}, booktitle = {Proceedings of the 35th International Conference on Machine Learning}, pages = {3150--3158}, publisher = {PMLR}, series = {Proceedings of Machine Learning Research}, title = {{Delayed Impact of Fair Machine Learning}}, url = {https://proceedings.mlr.press/v80/liu18c.html}, volume = {80}, year = {2018} } @misc{liu2023trustworthyllms, archiveprefix = {arXiv}, author = {Yang Liu and Yuanshun Yao and Jean-Francois Ton and Xiaoying Zhang and Ruocheng Guo and Hao Cheng and Yegor Klochkov and Muhammad Faaiz Taufiq and Hang Li}, eprint = {2308.05374}, note = {Version 2, revised March 2024}, primaryclass = {cs.CL}, title = {Trustworthy {LLMs}: A Survey and Guideline for Evaluating Large Language Models' Alignment}, url = {https://arxiv.org/abs/2308.05374}, year = {2023} } @misc{liu2026autoresearch, archiveprefix = {arXiv}, author = {Xiang Liu and Qumeng Sun and Bruno and Xiao Zhang and Jiahao Li and Haoyang Zhang}, eprint = {2608.17906}, note = {Preprint}, title = {{AutoResearch: Insight In, Hallucination Out}}, url = {https://arxiv.org/abs/2608.17906}, year = {2026} } @misc{liu2026evocowork, archiveprefix = {arXiv}, author = {Zhou Liu and Xingyan Chen and ZhaoYang Han and Zewei Pan and Zeli Su and Ningxin Pan and Xue Hu and Kaixiong Ting and Wentao Zhang}, eprint = {2610.14828}, title = {{EvoCowork: A Benchmark for Real-World Office Tasks}}, url = {https://arxiv.org/abs/2610.14828}, year = {2026} } @misc{liu2026feedbackdynamics, archiveprefix = {arXiv}, author = {Yuxuan Liu and Zhaochen Su and Yuhao Zhang and Jiahe Guo and Zhongwei Xie and Huihao Jing and Lingyun Xie and Qing Zong and Yauwai Yim and Zhixiong Zhang and Haoran Li and Yangqiu Song}, eprint = {2608.02636}, note = {Preprint}, title = {{Rethinking Self-Evolving Agent Skills: Feedback Dynamics over Multiple Rounds}}, url = {https://arxiv.org/abs/2608.02636}, year = {2026} } @misc{liu2026informationgain, archiveprefix = {arXiv}, author = {Wei Liu and Siya Qi and Yali Du and Yulan He}, eprint = {2603.02218}, note = {Preprint}, title = {{Self-Play Only Evolves When Self-Synthetic Pipeline Ensures Learnable Information Gain}}, url = {https://arxiv.org/abs/2603.02218}, year = {2026} } @inproceedings{liu2026shine, author = {Yewei Liu and Xiyuan Wang and Yansheng Mao and Yoav Gelberg and Haggai Maron and Muhan Zhang}, booktitle = {Proceedings of the 43rd International Conference on Machine Learning}, pages = {78777--78807}, publisher = {PMLR}, series = {Proceedings of Machine Learning Research}, title = {{SHINE: A Scalable In-Context Hypernetwork for Mapping Context to LoRA in a Single Pass}}, url = {https://proceedings.mlr.press/v306/liu26ek.html}, volume = {306}, year = {2026} } @inproceedings{liu2026whowhenpro, archiveprefix = {arXiv}, author = {Jiale Liu and Huajun Xi and Shaokun Zhang and Yifan Zeng and Tianwei Yue and Chi Wang and Jian Kang and Huazheng Wang and Qingyun Wu}, booktitle = {Advances in Neural Information Processing Systems}, eprint = {2607.09996}, note = {Accepted at NeurIPS 2026, Evaluations and Datasets Track; proceedings forthcoming}, title = {{Who\&When Pro: Can LLMs Really Attribute Failures in AI Agents?}}, url = {https://neurips.cc/virtual/2026/poster/139035}, year = {2026} } @misc{louck2026securingllmagentlongtermmemory, archiveprefix = {arXiv}, author = {Yedidel Louck}, eprint = {2606.24322}, howpublished = {arXiv preprint arXiv:2606.24322}, primaryclass = {cs.CR}, title = {{Securing LLM-Agent Long-Term Memory Against Poisoning: Non-Malleable, Origin-Bound Authority with Machine-Checked Guarantees}}, url = {https://arxiv.org/abs/2606.24322}, year = {2026} } @misc{lu2024aiscientist, archiveprefix = {arXiv}, author = {Chris Lu and Cong Lu and Robert Tjarko Lange and Jakob Foerster and Jeff Clune and David Ha}, eprint = {2408.06292}, note = {Preprint}, title = {{The AI Scientist: Towards Fully Automated Open-Ended Scientific Discovery}}, url = {https://arxiv.org/abs/2408.06292}, year = {2024} } @misc{lu2026verificationandselfimpro, archiveprefix = {arXiv}, author = {Chien-Ping Lu}, eprint = {2610.10611}, title = {{Verification and Self-Improvement in Agentic AI: Foundations and Limits}}, url = {https://arxiv.org/abs/2610.10611}, year = {2026} } @misc{luo2026harnessbank, archiveprefix = {arXiv}, author = {Xiaotian Luo and Dizhan Xue and Fengxingyu Wang and Chuanrui Hu and Yafeng Deng}, eprint = {2607.13683}, title = {{HarnessBank: Semantic Gene-Bank Search with Gated Verification for Agent-Harness Self-Evolution}}, url = {https://arxiv.org/abs/2607.13683}, year = {2026} } @misc{luo2026queryonlybackdoorattacksselfevolving, archiveprefix = {arXiv}, author = {Yuyang Luo and Haoran Wang and Kai Shu}, eprint = {2608.08303}, howpublished = {arXiv preprint arXiv:2608.08303}, primaryclass = {cs.AI}, title = {{Query-Only Backdoor Attacks on Self-Evolving Skills via Trajectory Poisoning}}, url = {https://arxiv.org/abs/2608.08303}, year = {2026} } @misc{lyu2026adam, archiveprefix = {arXiv}, author = {Xingyu Lyu and Jianfeng He and Ning Wang and Yidan Hu and Tao Li and Danjue Chen and Shixiong Li and Yimin Chen}, eprint = {2604.09747}, howpublished = {arXiv preprint arXiv:2604.09747}, primaryclass = {cs.CR}, title = {{ADAM: A Systematic Data Extraction Attack on Agent Memory via Adaptive Querying}}, url = {https://arxiv.org/abs/2604.09747}, year = {2026} } @inproceedings{ma2025falsifying, author = {Ma, George and Pfrommer, Samuel and Sojoudi, Somayeh}, booktitle = {Advances in Neural Information Processing Systems}, doi = {10.52202/085713-2122}, pages = {63175--63217}, publisher = {Curran Associates, Inc.}, title = {{Revising and Falsifying Sparse Autoencoder Feature Explanations}}, url = {https://proceedings.neurips.cc/paper_files/paper/2025/hash/5b6ce2529a79862a2d8ac57ac0e01d90-Abstract-Conference.html}, volume = {38, Main Conference}, year = {2025} } @inproceedings{ma2026dover, archiveprefix = {arXiv}, author = {Ma, Ming and Zhang, Jue and Yang, Fangkai and Kang, Yu and Lin, Qingwei and Rajmohan, Saravan and Zhang, Dongmei}, booktitle = {International Conference on Learning Representations}, eprint = {2512.06749}, note = {Revised January 2026}, pages = {28059--28087}, title = {{DoVer: Intervention-Driven Auto Debugging for LLM Multi-Agent Systems}}, url = {https://proceedings.iclr.cc/paper_files/paper/2026/hash/2f5fdce256ff0d7fb774da76e5e63209-Abstract-Conference.html}, volume = {2026}, year = {2026} } @misc{macdiarmid2025naturalemergent, archiveprefix = {arXiv}, author = {Monte MacDiarmid and Benjamin Wright and Jonathan Uesato and Joe Benton and Jon Kutasov and Sara Price and Naia Bouscal and Sam Bowman and Trenton Bricken and Alex Cloud and Carson Denison and Johannes Gasteiger and Ryan Greenblatt and Jan Leike and Jack Lindsey and Vlad Mikulik and Ethan Perez and Alex Rodrigues and Drake Thomas and Albert Webson and Daniel Ziegler and Evan Hubinger}, eprint = {2511.18397}, title = {{Natural Emergent Misalignment from Reward Hacking in Production RL}}, url = {https://arxiv.org/abs/2511.18397}, year = {2025} } @misc{madaan2023selfrefine, archiveprefix = {arXiv}, author = {Aman Madaan and Niket Tandon and Prakhar Gupta and Skyler Hallinan and Luyu Gao and Sarah Wiegreffe and Uri Alon and Nouha Dziri and Shrimai Prabhumoye and Yiming Yang and Shashank Gupta and Bodhisattwa Prasad Majumder and Katherine Hermann and Sean Welleck and Amir Yazdanbakhsh and Peter Clark}, eprint = {2303.17651}, title = {{Self-Refine: Iterative Refinement with Self-Feedback}}, url = {https://arxiv.org/abs/2303.17651}, year = {2023} } @inproceedings{makelov2024illusion, author = {Makelov, Aleksandar and Lange, Georg and Geiger, Atticus and Nanda, Neel}, booktitle = {International Conference on Learning Representations}, pages = {26486--26515}, title = {{Is This the Subspace You Are Looking for? An Interpretability Illusion for Subspace Activation Patching}}, url = {https://proceedings.iclr.cc/paper_files/paper/2024/hash/70b8505ac79e3e131756f793cd80eb8d-Abstract-Conference.html}, volume = {2024}, year = {2024} } @inproceedings{malladi2026opex, author = {Malladi, Advaith and Srivastava, Shashank}, booktitle = {Proceedings of the 43rd International Conference on Machine Learning}, pages = {86027--86043}, publisher = {PMLR}, title = {{Directly Optimizing Natural Language Explanations for Behavioral Faithfulness: Simulatability and Recoverability}}, url = {https://proceedings.mlr.press/v306/malladi26a.html}, volume = {306}, year = {2026} } @misc{mansuri2026doc2lora, archiveprefix = {arXiv}, author = {Chand Sahil Mansuri and Joel Zachariah and Sadamori Kojaku}, eprint = {2609.38374}, title = {{Doc2LoRA Provides Decodable Representations of Scientific Ideas}}, url = {https://arxiv.org/abs/2609.38374}, year = {2026} } @misc{mao2026agentsremember, archiveprefix = {arXiv}, author = {Xutao Mao and Liangjie Zhao and Leyao Wang and Rui Qian and Qiang Huang and Wentao Wang and Bo Han and Xiang Zheng and Cong Wang}, eprint = {2607.10526}, title = {{Agents Don't Just Agree, They Remember: Benchmarking Persistent Sycophancy in Self-Improving Personal Agents}}, url = {https://arxiv.org/abs/2607.10526}, year = {2026} } @misc{mao2026safeevolveharnesspolicycoevolutionagent, archiveprefix = {arXiv}, author = {Qinghua Mao and Wanying Qu and Dadi Guo and Leitao Yuan and Qingyu Liu and Yu Li and Guanxu Chen and Yanwei Fu and Xi Lin and Xia Hu and Dongrui Liu}, eprint = {2609.02786}, primaryclass = {cs.AI}, title = {SafeEvolve: Harness-Policy Co-Evolution from Agent Experience for Safety Alignment}, url = {https://arxiv.org/abs/2609.02786}, year = {2026} } @inproceedings{marks2025circuits, author = {Marks, Samuel and Rager, Can and Michaud, Eric and Belinkov, Yonatan and Bau, David and Mueller, Aaron}, booktitle = {International Conference on Learning Representations}, pages = {23888--23923}, title = {{Sparse Feature Circuits: Discovering and Editing Interpretable Causal Graphs in Language Models}}, url = {https://proceedings.iclr.cc/paper_files/paper/2025/hash/3ba4d47a83e498c2b1a0868cba20f6de-Abstract-Conference.html}, volume = {2025}, year = {2025} } @inproceedings{mayne2026positivefaithfulness, author = {Mayne, Harry and Kang, Justin Singh and Gould, Dewi Sid William and Ramchandran, Kannan and Mahdi, Adam and Siegel, Noah Y.}, booktitle = {Proceedings of the 43rd International Conference on Machine Learning}, pages = {87135--87185}, publisher = {PMLR}, title = {{A Positive Case for Faithfulness: {LLM} Self-Explanations Help Predict Model Behavior}}, url = {https://proceedings.mlr.press/v306/mayne26a.html}, volume = {306}, year = {2026} } @inproceedings{mckenzie2026steeringresistance, author = {Mckenzie, Alex and Pepper, Keenan and Servaes, Stijn and Leitgab, Martin and Cubuktepe, Murat and Vaiana, Michael and De Lucena, Diogo S and Rosenblatt, Judd and Graziano, Michael S. A.}, booktitle = {Proceedings of the 43rd International Conference on Machine Learning}, pages = {87297--87324}, publisher = {PMLR}, title = {{Endogenous Resistance to Activation Steering in Language Models}}, url = {https://proceedings.mlr.press/v306/mckenzie26a.html}, volume = {306}, year = {2026} } @misc{mecca2026skillsecurerdetectingpatchingpromptinjection, archiveprefix = {arXiv}, author = {Donato Mecca and Alberto Verna and Youness Bouchari and Nikhil Jha and Marco Mellia}, eprint = {2609.14079}, howpublished = {arXiv preprint arXiv:2609.14079}, primaryclass = {cs.CR}, title = {{SkillSecurer: Detecting and Patching Prompt-Injection Vulnerabilities in AI Agent Skills}}, url = {https://arxiv.org/abs/2609.14079}, year = {2026} } @inproceedings{meloux2025identifiability, author = {M\'{e}loux, Maxime and Maniu, Silviu and Portet, Fran\c{c}ois and Peyrard, Maxime}, booktitle = {International Conference on Learning Representations}, pages = {32900--32923}, title = {{Everything, Everywhere, All at Once: Is Mechanistic Interpretability Identifiable?}}, url = {https://proceedings.iclr.cc/paper_files/paper/2025/hash/51157713f1f1f572a05647e1aef1b1ed-Abstract-Conference.html}, volume = {2025}, year = {2025} } @misc{meng2026rsibench, archiveprefix = {arXiv}, author = {Meng, Fanqing and Du, Lingxiao and Chen, Qiguang and Zhao, Ziqi and Lu, Haocheng and Hu, Mengkang and Shieh, Michael Qizhe}, eprint = {2607.25886}, title = {{RSIBench-Data: Benchmarking Data-Centric Research for Recursive Self-Improvement}}, url = {https://arxiv.org/abs/2607.25886}, year = {2026} } @inproceedings{mickel2025representation, author = {Jennifer Mickel and Maria De-Arteaga and Liu Leqi and Kevin Tian}, booktitle = {Advances in Neural Information Processing Systems}, title = {{More of the Same: Persistent Representational Harms Under Increased Representation}}, url = {https://proceedings.neurips.cc/paper_files/paper/2025/hash/584c609e4d9822eafa0304e8757ed8c2-Abstract-Conference.html}, volume = {38}, year = {2025} } @misc{mireshghallah2025cimemoriescompositionalbenchmarkcontextual, archiveprefix = {arXiv}, author = {Niloofar Mireshghallah and Neal Mangaokar and Narine Kokhlikyan and Arman Zharmagambetov and Manzil Zaheer and Saeed Mahloujifar and Kamalika Chaudhuri}, eprint = {2511.14937}, howpublished = {arXiv preprint arXiv:2511.14937}, primaryclass = {cs.CR}, title = {{CIMemories: A Compositional Benchmark for Contextual Integrity of Persistent Memory in LLMs}}, url = {https://arxiv.org/abs/2511.14937}, year = {2025} } @misc{mishra2026prism, archiveprefix = {arXiv}, author = {Vaibhav Mishra}, eprint = {2603.13309}, note = {Preprint}, title = {{Preventing Curriculum Collapse in Self-Evolving Reasoning Systems}}, url = {https://arxiv.org/abs/2603.13309}, year = {2026} } @techreport{mitchell1980bias, author = {Tom M. Mitchell}, institution = {Department of Computer Science, Rutgers University}, number = {CBM-TR-117}, title = {The Need for Biases in Learning Generalizations}, year = {1980} } @book{mitchell1997ml, author = {Tom M. Mitchell}, publisher = {McGraw-Hill}, title = {Machine Learning}, year = {1997} } @inproceedings{mitchell2019modelcards, author = {Mitchell, Margaret and Wu, Simone and Zaldivar, Andrew and Barnes, Parker and Vasserman, Lucy and Hutchinson, Ben and Spitzer, Elena and Raji, Inioluwa Deborah and Gebru, Timnit}, booktitle = {Proceedings of the Conference on Fairness, Accountability, and Transparency}, doi = {10.1145/3287560.3287596}, pages = {220--229}, title = {{Model Cards for Model Reporting}}, url = {https://arxiv.org/abs/1810.03993}, year = {2019} } @misc{moll2026grasp, archiveprefix = {arXiv}, author = {Johannes Moll and Jean-Philippe Corbeil and Jiazhen Pan and Martin Hadamitzky and Daniel Rueckert and Lisa Adams and Keno Bressem}, eprint = {2605.29668}, note = {Preprint}, title = {{GRASP: Gated Regression-Aware Skill Proposer for Self-Improving LLM Agents}}, url = {https://arxiv.org/abs/2605.29668}, year = {2026} } @inproceedings{motwani2024collusion, author = {Motwani, Sumeet Ramesh and Baranchuk, Mikhail and Strohmeier, Martin and Bolina, Vijay and Torr, Philip and Hammond, Lewis and de Witt, Christian Schroeder}, booktitle = {Advances in Neural Information Processing Systems}, doi = {10.52202/079017-2336}, pages = {73439--73486}, publisher = {Curran Associates, Inc.}, title = {{Secret Collusion among AI Agents: Multi-Agent Deception via Steganography}}, url = {https://proceedings.neurips.cc/paper_files/paper/2024/hash/861f7dad098aec1c3560fb7add468d41-Abstract-Conference.html}, volume = {37}, year = {2024} } @misc{mouret2015mapelites, archiveprefix = {arXiv}, author = {Jean-Baptiste Mouret and Jeff Clune}, eprint = {1504.04909}, note = {Preprint}, title = {{Illuminating search spaces by mapping elites}}, url = {https://arxiv.org/abs/1504.04909}, year = {2015} } @inproceedings{mueller2025mib, author = {Mueller, Aaron and Geiger, Atticus and Wiegreffe, Sarah and Arad, Dana and Arcuschin, Iv\'{a}n and Belfki, Adam and Chan, Yik Siu and Fiotto-Kaufman, Jaden Fried and Haklay, Tal and Hanna, Michael and Huang, Jing and Gupta, Rohan and Nikankin, Yaniv and Orgad, Hadas and Prakash, Nikhil and Reusch, Anja and Sankaranarayanan, Aruna and Shao, Shun and Stolfo, Alessandro and Tutek, Martin and Zur, Amir and Bau, David and Belinkov, Yonatan}, booktitle = {Proceedings of the 42nd International Conference on Machine Learning}, pages = {45069--45108}, publisher = {PMLR}, title = {{{MIB}: A Mechanistic Interpretability Benchmark}}, url = {https://proceedings.mlr.press/v267/mueller25a.html}, volume = {267}, year = {2025} } @inproceedings{natarajan2026targetedprobes, author = {Natarajan, Vikram and Jain, Devina and Arora, Shivam and Golechha, Satvik and Bloom, Joseph Isaac}, booktitle = {Proceedings of the 43rd International Conference on Machine Learning}, pages = {91932--91947}, publisher = {PMLR}, title = {{One Probe Won't Catch Them All: Towards Targeted Deception Detection}}, url = {https://proceedings.mlr.press/v306/natarajan26a.html}, volume = {306}, year = {2026} } @inproceedings{nguyen2026pidsteering, author = {Nguyen, Dung Viet and Pham, Yen and Vu, Hieu and Zhang, Lei and Nguyen, Tan}, booktitle = {International Conference on Learning Representations}, pages = {154471--154505}, title = {{Activation Steering with a Feedback Controller}}, url = {https://proceedings.iclr.cc/paper_files/paper/2026/hash/fa5617c176e76fee83f3f9947fdf9f3f-Abstract-Conference.html}, volume = {2026}, year = {2026} } @misc{nguyen2026rsea, archiveprefix = {arXiv}, author = {Michael Nguyen and Quoc Nguyen and Paul Vuong}, eprint = {2606.28374}, note = {Preprint}, title = {{Recursive Self-Evolving Agents via Held-Out Selection}}, url = {https://arxiv.org/abs/2606.28374}, year = {2026} } @misc{nian2026auditable, archiveprefix = {arXiv}, author = {Nian, Yi and Yuan, Aojie and Zhang, Haiyue and Li, Jiate and Li, Li and Hu, Xiyang and Wei, Hua and Xiao, Xiongye and Xiao, Chaowei and Zhao, Yue}, eprint = {2604.05485}, title = {{Auditable Agents}}, url = {https://arxiv.org/abs/2604.05485}, year = {2026} } @techreport{nist2023airmf, author = {{National Institute of Standards and Technology}}, doi = {10.6028/NIST.AI.100-1}, institution = {National Institute of Standards and Technology}, number = {NIST AI 100-1}, title = {Artificial Intelligence Risk Management Framework ({AI RMF} 1.0)}, url = {https://doi.org/10.6028/NIST.AI.100-1}, year = {2023} } @misc{novikov2025alphaevolve, archiveprefix = {arXiv}, author = {Alexander Novikov and Ngân Vũ and Marvin Eisenberger and Emilien Dupont and Po-Sen Huang and Adam Zsolt Wagner and Sergey Shirobokov and Borislav Kozlovskii and Francisco J. R. Ruiz and Abbas Mehrabian and M. Pawan Kumar and Abigail See and Swarat Chaudhuri and George Holland and Alex Davies and Sebastian Nowozin and Pushmeet Kohli and Matej Balog}, eprint = {2506.13131}, note = {Preprint}, title = {{AlphaEvolve: A coding agent for scientific and algorithmic discovery}}, url = {https://arxiv.org/abs/2506.13131}, year = {2025} } @misc{openai2026gpt6astra, author = {{OpenAI}}, howpublished = {OpenAI}, month = {September}, title = {{GPT-6 Astra}: A New Generation of Intelligence}, url = {https://openai.com/index/gpt-6-astra/}, year = {2026} } @inproceedings{orseau2016interruptible, author = {Orseau, Laurent and Armstrong, Stuart}, booktitle = {Proceedings of the Thirty-Second Conference on Uncertainty in Artificial Intelligence (UAI 2016)}, title = {{Safely Interruptible Agents}}, url = {https://auai.org/~w-auai/uai2016/proceedings/papers/68.pdf}, year = {2016} } @inproceedings{ouyang2026reasoningbank, author = {Siru Ouyang and Jun Yan and I-Hung Hsu and Yanfei Chen and Ke Jiang and Zifeng Wang and Rujun Han and Long T. Le and Samira Daruki and Xiangru Tang and Vishy Tirumalashetty and George Lee and Mahsan Rofouei and Hangfei Lin and Jiawei Han and Chen-Yu Lee and Tomas Pfister}, booktitle = {International Conference on Learning Representations}, title = {{ReasoningBank}: Scaling Agent Self-Evolving with Reasoning Memory}, url = {https://openreview.net/forum?id=jL7fwchScm}, year = {2026} } @inproceedings{overman2026conservatism, author = {Overman, William and Bayati, Mohsen}, booktitle = {Proceedings of the 43rd International Conference on Machine Learning}, pages = {95349--95379}, publisher = {PMLR}, title = {{Calibrating Conservatism for Scalable Oversight}}, url = {https://proceedings.mlr.press/v306/overman26a.html}, volume = {306}, year = {2026} } @inproceedings{overman2026oversightgame, author = {Overman, William and Bayati, Mohsen}, booktitle = {Proceedings of the 43rd International Conference on Machine Learning}, pages = {95380--95413}, publisher = {PMLR}, title = {{The Oversight Game: Learning to Cooperatively Balance an {AI} Agent's Safety and Autonomy}}, url = {https://proceedings.mlr.press/v306/overman26b.html}, volume = {306}, year = {2026} } @misc{pan2026theworldwontwait, archiveprefix = {arXiv}, author = {Ningxin Pan and Hanyu Li and Yeqing Shen and Ning Ding and Yehui Tang and Wentao Zhang}, eprint = {2610.20665}, title = {{The World Won't Wait: Evaluating and Teaching Agents Under a Live Clock}}, url = {https://arxiv.org/abs/2610.20665}, year = {2026} } @misc{pathmanathan2026compliance, archiveprefix = {arXiv}, author = {Pankayaraj Pathmanathan and Furong Huang}, eprint = {2607.27594}, title = {{Compliance2LoRA: Personalizable On-Demand Safety Alignment on Arbitrary Policy Subsets via Hypernetwork-Generated LoRA Adapters}}, url = {https://arxiv.org/abs/2607.27594}, year = {2026} } @inproceedings{peng2025regenesis, author = {Peng, Xiangyu and Xia, Congying and Yang, Xinyi and Xiong, Caiming and Wu, Chien-Sheng and Xing, Chen}, booktitle = {International Conference on Learning Representations}, pages = {62484--62519}, title = {{ReGenesis}: {LLMs} can Grow into Reasoning Generalists via Self-Improvement}, url = {https://proceedings.iclr.cc/paper_files/paper/2025/hash/9c77f2ce42151b2c2e26d2cf47f99564-Abstract-Conference.html}, year = {2025} } @misc{perez2022modelwritten, archiveprefix = {arXiv}, author = {Ethan Perez and Sam Ringer and Kamilė Lukošiūtė and Karina Nguyen and Edwin Chen and Scott Heiner and Craig Pettit and Catherine Olsson and Sandipan Kundu and Saurav Kadavath and Andy Jones and Anna Chen and Ben Mann and Brian Israel and Bryan Seethor and Cameron McKinnon and Christopher Olah and Da Yan and Daniela Amodei and Dario Amodei and Dawn Drain and Dustin Li and Eli Tran-Johnson and Guro Khundadze and Jackson Kernion and James Landis and Jamie Kerr and Jared Mueller and Jeeyoon Hyun and Joshua Landau and Kamal Ndousse and Landon Goldberg and Liane Lovitt and Martin Lucas and Michael Sellitto and Miranda Zhang and Neerav Kingsland and Nelson Elhage and Nicholas Joseph and Noemí Mercado and Nova DasSarma and Oliver Rausch and Robin Larson and Sam McCandlish and Scott Johnston and Shauna Kravec and Sheer El Showk and Tamera Lanham and Timothy Telleen-Lawton and Tom Brown and Tom Henighan and Tristan Hume and Yuntao Bai and Zac Hatfield-Dodds and Jack Clark and Samuel R. Bowman and Amanda Askell and Roger Grosse and Danny Hernandez and Deep Ganguli and Evan Hubinger and Nicholas Schiefer and Jared Kaplan}, eprint = {2212.09251}, note = {Preprint}, title = {{Discovering Language Model Behaviors with Model-Written Evaluations}}, url = {https://arxiv.org/abs/2212.09251}, year = {2022} } @inproceedings{ponkshe2026safetysubspaces, author = {Kaustubh Ponkshe and Shaan Shah and Raghav Singhal and Praneeth Vepakomma}, booktitle = {International Conference on Learning Representations}, title = {{Safety Subspaces are Not Linearly Distinct: A Fine-Tuning Case Study}}, url = {https://proceedings.iclr.cc/paper_files/paper/2026/hash/9a9f4e15ad0d680429a3e0570a96f763-Abstract-Conference.html}, year = {2026} } @inproceedings{prasad2025scpo, author = {Archiki Prasad and Weizhe Yuan and Richard Yuanzhe Pang and Jing Xu and Maryam Fazel-Zarandi and Mohit Bansal and Sainbayar Sukhbaatar and Jason E Weston and Jane Yu}, booktitle = {Proceedings of the 42nd International Conference on Machine Learning}, pages = {49737--49751}, publisher = {PMLR}, series = {Proceedings of Machine Learning Research}, title = {{Self-Consistency Preference Optimization}}, url = {https://proceedings.mlr.press/v267/prasad25a.html}, volume = {267}, year = {2025} } @inproceedings{prinster2025watch, author = {Prinster, Drew and Han, Xing and Liu, Anqi and Saria, Suchi}, booktitle = {Proceedings of the 42nd International Conference on Machine Learning}, pages = {49830--49859}, publisher = {PMLR}, title = {{{WATCH}: Adaptive Monitoring for {AI} Deployments via Weighted-Conformal Martingales}}, url = {https://proceedings.mlr.press/v267/prinster25a.html}, volume = {267}, year = {2025} } @article{pugh2016qualitydiversity, author = {Justin K. Pugh and Lisa B. Soros and Kenneth O. Stanley}, doi = {10.3389/frobt.2016.00040}, journal = {Frontiers in Robotics and AI}, pages = {40}, title = {{Quality Diversity: A New Frontier for Evolutionary Computation}}, url = {https://www.frontiersin.org/journals/robotics-and-ai/articles/10.3389/frobt.2016.00040/full}, volume = {3}, year = {2016} } @inproceedings{pulipaka2026persistbench, archiveprefix = {arXiv}, author = {Sidharth Pulipaka and Oliver Chen and Manas Sharma and Taaha S Bajwa and Vyas Raina and Ivaxi Sheth}, booktitle = {International Conference on Machine Learning}, eprint = {2602.01146}, title = {{PersistBench: When Should Long-Term Memories Be Forgotten by LLMs?}}, url = {https://arxiv.org/abs/2602.01146}, year = {2026} } @inproceedings{qi2024finetuning, author = {Qi, Xiangyu and Zeng, Yi and Xie, Tinghao and Chen, Pin-Yu and Jia, Ruoxi and Mittal, Prateek and Henderson, Peter}, booktitle = {International Conference on Learning Representations}, title = {Fine-Tuning Aligned Language Models Compromises Safety, Even When Users Do Not Intend To!}, url = {https://openreview.net/forum?id=hTEGyKf0dZ}, year = {2024} } @inproceedings{qi2025durability, author = {Xiangyu Qi and Boyi Wei and Nicholas Carlini and Yangsibo Huang and Tinghao Xie and Luxi He and Matthew Jagielski and Milad Nasr and Prateek Mittal and Peter Henderson}, booktitle = {International Conference on Learning Representations}, title = {{On Evaluating the Durability of Safeguards for Open-Weight LLMs}}, url = {https://proceedings.iclr.cc/paper_files/paper/2025/hash/9d3a4cdf6f70559e8c6fe02170fba568-Abstract-Conference.html}, year = {2025} } @misc{qi2026generalizationgap, archiveprefix = {arXiv}, author = {Zhenting Qi and Susanna Maria Baby and Stefanie Anna Baby and Kan Yuan and Andrew Tomkins and Tu Vu and Da-Cheng Juan and Cyrus Rashtchian}, eprint = {2606.01075}, title = {{On the Generalization Gap in Self-Evolving Language Model Reasoning}}, url = {https://arxiv.org/abs/2606.01075}, year = {2026} } @article{qi2026trustworthyagentic, author = {Qi, Jinhu and Li, Muzhi and Liu, Jiahong and Shu, Yuqin and Yu, Dianzhi and Ma, Shicheng and Cui, Wenqian and Zhao, Yiyang and Chen, Yiyi and Jiang, Ruoxi and King, Irwin and Xu, Zenglin}, doi = {10.20935/AcadAI8260}, journal = {Academia AI and Applications}, title = {Towards trustworthy agentic AI: a comprehensive survey of safety, robustness, privacy, and system security}, url = {https://arxiv.org/abs/2605.23989}, volume = {2}, year = {2026} } @misc{qu2026she, archiveprefix = {arXiv}, author = {Wanying Qu and Qinghua Mao and Yu Li and Jiyao Liu and Xin Zhang and Dadi Guo and Yanxu Zhu and Qingyu Liu and Leitao Yuan and Xi Lin and Shanfeng Zhu and Yanwei Fu and Jing Shao and Xia Hu and Dongrui Liu}, eprint = {2608.09885}, howpublished = {arXiv preprint arXiv:2608.09885}, title = {{SHE: Trajectory-driven Safety Harness Evolution for LLM Agents}}, url = {https://arxiv.org/abs/2608.09885}, year = {2026} } @inproceedings{rahman2025humandebate, author = {Rahman, Salman and Issaka, Sheriff and Suvarna, Ashima and Liu, Genglin and Shiffer, James and Lee, Jaeyoung and Parvez, Md Rizwan and Palangi, Hamid and Feng, Shi and Peng, Nanyun and Choi, Yejin and Michael, Julian and Jiang, Liwei and Gabriel, Saadia}, booktitle = {Advances in Neural Information Processing Systems}, doi = {10.52202/085713-5666}, pages = {170218--170297}, publisher = {Curran Associates, Inc.}, title = {{AI Debate Aids Assessment of Controversial Claims}}, url = {https://proceedings.neurips.cc/paper_files/paper/2025/hash/f88cc8930b47a45ec4733123bf3039b9-Abstract-Conference.html}, volume = {38, Main Conference}, year = {2025} } @misc{raji2020accountability, archiveprefix = {arXiv}, author = {Raji, Inioluwa Deborah and Smart, Andrew and White, Rebecca N. and Mitchell, Margaret and Gebru, Timnit and Hutchinson, Ben and Smith-Loud, Jamila and Theron, Daniel and Barnes, Parker}, eprint = {2001.00973}, title = {{Closing the AI Accountability Gap: Defining an End-to-End Framework for Internal Algorithmic Auditing}}, url = {https://arxiv.org/abs/2001.00973}, year = {2020} } @misc{rank2026posttrainbench, archiveprefix = {arXiv}, author = {Rank, Ben and Bhatnagar, Hardik and Prabhu, Ameya and Eisenberg, Shira and Nguyen, Karina and Bethge, Matthias and Andriushchenko, Maksym}, eprint = {2603.08640}, title = {{PostTrainBench: Can LLM Agents Automate LLM Post-Training?}}, url = {https://arxiv.org/abs/2603.08640}, year = {2026} } @misc{real2020automlzero, archiveprefix = {arXiv}, author = {Esteban Real and Chen Liang and David R. So and Quoc V. Le}, eprint = {2003.03384}, title = {{AutoML-Zero: Evolving Machine Learning Algorithms From Scratch}}, url = {https://arxiv.org/abs/2003.03384}, year = {2020} } @misc{ren2026selfimprovements, archiveprefix = {arXiv}, author = {Ren, Zhe and Chen, Yimeng and Guo, Dandan and Rong, Guowei and Li, Tonghui and Xiong, R. B. and Lan, Qingfeng and Wang, Wenyi and Li, Nanbo and Yang, Yibo and Zhuge, Mingchen and Schmidhuber, J{\"u}rgen}, eprint = {2607.13104}, primaryclass = {cs.AI}, title = {Self-Improvements in Modern Agentic Systems: A Survey}, url = {https://arxiv.org/abs/2607.13104}, year = {2026} } @inproceedings{richter2025behavioralshift, author = {Richter, Leo and He, Xuanli and Minervini, Pasquale and Kusner, Matt}, booktitle = {International Conference on Learning Representations}, pages = {49313--49339}, title = {{An Auditing Test to Detect Behavioral Shift in Language Models}}, url = {https://proceedings.iclr.cc/paper_files/paper/2025/hash/7af140f516c9f57050f7359bf53bc8f0-Abstract-Conference.html}, volume = {2025}, year = {2025} } @misc{robeyns2025sica, archiveprefix = {arXiv}, author = {Maxime Robeyns and Martin Szummer and Laurence Aitchison}, eprint = {2504.15228}, note = {Preprint}, title = {{A Self-Improving Coding Agent}}, url = {https://arxiv.org/abs/2504.15228}, year = {2025} } @misc{roesner2026reflectionstrustingtrustrevisited, archiveprefix = {arXiv}, author = {Franziska Roesner and Tadayoshi Kohno}, eprint = {2609.17817}, primaryclass = {cs.CR}, title = {Reflections on Trusting Trust, Revisited: Contaminating Self-Modifying AI Coding Agents with Poisoned Benchmarks}, url = {https://arxiv.org/abs/2609.17817}, year = {2026} } @misc{roesner2026trustingtrust, archiveprefix = {arXiv}, author = {Franziska Roesner and Tadayoshi Kohno}, eprint = {2609.17817}, title = {{Reflections on Trusting Trust, Revisited: Contaminating Self-Modifying AI Coding Agents with Poisoned Benchmarks}}, url = {https://arxiv.org/abs/2609.17817}, year = {2026} } @misc{roger2023benchmarksdetectingmeasurementtampering, archiveprefix = {arXiv}, author = {Fabien Roger and Ryan Greenblatt and Max Nadeau and Buck Shlegeris and Nate Thomas}, eprint = {2308.15605}, howpublished = {arXiv preprint arXiv:2308.15605}, primaryclass = {cs.LG}, title = {{Benchmarks for Detecting Measurement Tampering}}, url = {https://arxiv.org/abs/2308.15605}, year = {2023} } @inproceedings{rosser2025agentbreeder, author = {J Rosser and Jakob Foerster}, booktitle = {Advances in Neural Information Processing Systems}, doi = {10.52202/085713-5003}, title = {{AgentBreeder}: Mitigating the {AI} Safety Risks of Multi-Agent Scaffolds via Self-Improvement}, url = {https://proceedings.nips.cc/paper_files/paper/2025/hash/dc2ccde7ee43e5719e08c68e848bd65a-Abstract-Conference.html}, volume = {38}, year = {2025} } @misc{roth2026hackverifiable, archiveprefix = {arXiv}, author = {Amit Roth and Ankur Samanta and Matan Halevy and Yoav Levine and Yonathan Efroni}, eprint = {2605.20744}, note = {Preprint}, title = {{Hack-Verifiable Environments: Towards Evaluating Reward Hacking at Scale}}, url = {https://arxiv.org/abs/2605.20744}, year = {2026} } @inproceedings{roytburg2026legibility, archiveprefix = {arXiv}, author = {Dani Roytburg and Shreya Sridhar and Daphne Ippolito}, booktitle = {Advances in Neural Information Processing Systems}, eprint = {2603.20508}, note = {Accepted at NeurIPS 2026; proceedings forthcoming}, title = {{Measuring Weak-to-Strong Legibility of Reasoning Models}}, url = {https://neurips.cc/virtual/2026/poster/153826}, year = {2026} } @inproceedings{ryd2026removingsandbagging, author = {Ryd, Emil and Bartsch, Henning and Stastny, Julian and Benton, Joe and Hebbar, Vivek}, booktitle = {Proceedings of the 43rd International Conference on Machine Learning}, pages = {106456--106500}, publisher = {PMLR}, title = {{Removing Sandbagging in {LLM}s by Training with Weak Supervision}}, url = {https://proceedings.mlr.press/v306/ryd26a.html}, volume = {306}, year = {2026} } @misc{sah2026evoundo, archiveprefix = {arXiv}, author = {Tanmay Sah and Dolly Sah and Harshul Jain and Tanya Sah}, eprint = {2608.28363}, primaryclass = {cs.AI}, title = {{EvoUndo}: Recoverability-Constrained Self-Evolution for {LLM} Agent Harnesses}, url = {https://arxiv.org/abs/2608.28363}, year = {2026} } @misc{sahoo2026sahoo, archiveprefix = {arXiv}, author = {Subramanyam Sahoo and Aman Chadha and Vinija Jain and Divya Chaudhary}, eprint = {2603.06333}, howpublished = {arXiv preprint arXiv:2603.06333}, title = {{SAHOO: Safeguarded Alignment for High-Order Optimization Objectives in Recursive Self-Improvement}}, url = {https://arxiv.org/abs/2603.06333}, year = {2026} } @inproceedings{samanta2026thoughtics, author = {Samanta, Ankur and Magesh, Akshayaa and Jain, Ayush and Asadi, Kavosh and Yu, Youliang and Jiang, Daniel R. and Vidolov, Boris and Hassani, Kaveh and Sajda, Paul and Bhandari, Jalaj and Efroni, Yonathan}, booktitle = {Proceedings of the 43rd International Conference on Machine Learning}, pages = {107412--107440}, publisher = {PMLR}, title = {{Structure Enables Effective Self-Localization of Errors in {LLM}s}}, url = {https://proceedings.mlr.press/v306/samanta26a.html}, volume = {306}, year = {2026} } @misc{schmidgall2025agentrxiv, archiveprefix = {arXiv}, author = {Samuel Schmidgall and Michael Moor}, eprint = {2503.18102}, note = {Preprint}, title = {{AgentRxiv: Towards Collaborative Autonomous Research}}, url = {https://arxiv.org/abs/2503.18102}, year = {2025} } @misc{schmidhuber2006godel, archiveprefix = {arXiv}, author = {Juergen Schmidhuber}, eprint = {cs/0309048}, title = {{Goedel Machines: Self-Referential Universal Problem Solvers Making Provably Optimal Self-Improvements}}, url = {https://arxiv.org/abs/cs/0309048}, year = {2006} } @misc{schmidhuber2011powerplay, archiveprefix = {arXiv}, author = {Jürgen Schmidhuber}, eprint = {1112.5309}, note = {Preprint}, title = {{POWERPLAY: Training an Increasingly General Problem Solver by Continually Searching for the Simplest Still Unsolvable Problem}}, url = {https://arxiv.org/abs/1112.5309}, year = {2011} } @book{schumacher1973small, address = {London}, author = {Schumacher, E. F.}, publisher = {Blond \& Briggs}, title = {Small Is Beautiful: Economics as if People Mattered}, year = {1973} } @misc{sengupta2026certificates, archiveprefix = {arXiv}, author = {Biswa Sengupta}, eprint = {2607.00871}, note = {Preprint}, title = {{Self-Evolving Agents with Anytime-Valid Certificates}}, url = {https://arxiv.org/abs/2607.00871}, year = {2026} } @misc{seo2025factverifiers, archiveprefix = {arXiv}, author = {Wooseok Seo and Seungju Han and Jaehun Jung and Benjamin Newman and Seungwon Lim and Seungbeen Lee and Ximing Lu and Yejin Choi and Youngjae Yu}, eprint = {2506.13342}, note = {Preprint}, title = {{Verifying the Verifiers: Unveiling Pitfalls and Potentials in Fact Verifiers}}, url = {https://arxiv.org/abs/2506.13342}, year = {2025} } @misc{shang2026precommit, archiveprefix = {arXiv}, author = {Shang, Linfang and Xu, Ming and Sun, Yiding and Xia, Tianle and Hu, Lingxiang and Xu, Lan and Zheng, Ning}, eprint = {2608.05810}, note = {Preprint}, title = {When Self-Evolution Backfires: Pre-Commit Gating against Skill Contamination in {LLM} Agents}, url = {https://arxiv.org/abs/2608.05810}, year = {2026} } @inproceedings{shao2024privacylens, author = {Shao, Yijia and Li, Tianshi and Shi, Weiyan and Liu, Yanchen and Yang, Diyi}, booktitle = {Advances in Neural Information Processing Systems}, doi = {10.52202/079017-2837}, editor = {A. Globerson and L. Mackey and D. Belgrave and A. Fan and U. Paquet and J. Tomczak and C. Zhang}, pages = {89373--89407}, publisher = {Curran Associates, Inc.}, title = {{PrivacyLens: Evaluating Privacy Norm Awareness of Language Models in Action}}, url = {https://proceedings.neurips.cc/paper_files/paper/2024/file/a2a7e58309d5190082390ff10ff3b2b8-Paper-Datasets_and_Benchmarks_Track.pdf}, volume = {37}, year = {2024} } @inproceedings{shao2026misevolve, author = {Shuai Shao and Qihan Ren and Dongrui Liu and Chen Qian and Boyi Wei and Dadi Guo and Jingyi Yang and Xinhao Song and Linfeng Zhang and Weinan Zhang and Jing Shao}, booktitle = {International Conference on Learning Representations}, title = {Your Agent May Misevolve: Emergent Risks in Self-evolving {LLM} Agents}, url = {https://openreview.net/forum?id=Fd1jgQQW28}, year = {2026} } @misc{shao2026protocolvalidity, archiveprefix = {arXiv}, author = {Jiaqi Shao and Hanck Chen and Wei Zhang and Maxm Pan and Bing Luo}, eprint = {2607.22368}, note = {Preprint}, title = {{Do Agent Benchmarks Measure Capability? Protocol Validity in the Age of Agentic AI}}, url = {https://arxiv.org/abs/2607.22368}, year = {2026} } @inproceedings{shapira2026sycophancy, author = {Shapira, Itai and Benade, Gerdus and Procaccia, Ariel D.}, booktitle = {Proceedings of the 43rd International Conference on Machine Learning}, pages = {110417--110441}, publisher = {PMLR}, series = {Proceedings of Machine Learning Research}, title = {{How RLHF Amplifies Sycophancy}}, url = {https://proceedings.mlr.press/v306/shapira26a.html}, volume = {306}, year = {2026} } @misc{shawn2026pace, archiveprefix = {arXiv}, author = {Shawn, Zayx}, eprint = {2606.08106}, title = {{PACE: Anytime-Valid Acceptance Tests for Self-Evolving Agents}}, url = {https://arxiv.org/abs/2606.08106}, year = {2026} } @article{shen2026controllable, author = {Chenglei Shen and Xiao Zhang and Teng Shi and Changshuo Zhang and Guofu Xie and Jun Xu and Ming He and Jianping Fan}, doi = {10.1007/s11704-025-41366-5}, journal = {Frontiers of Computer Science}, number = {10}, pages = {2010619}, title = {{A Survey of controllable learning: {M}ethods and applications in information retrieval}}, url = {https://journal.hep.com.cn/fcs/EN/10.1007/s11704-025-41366-5}, volume = {20}, year = {2026} } @inproceedings{shen2026stylitruth, author = {Shen, Chenglei and Sun, Zhongxiang and Shi, Teng and Zhang, Xiao and Xu, Jun}, booktitle = {International Conference on Learning Representations}, pages = {2531--2547}, title = {StyliTruth: Unlocking Stylized yet Truthful LLM Generation via Disentangled Steering}, volume = {2026}, year = {2026} } @inproceedings{sheth2026openended, author = {Sheth, Ivaxi and Wehner, Jan and Abdelnabi, Sahar and Binkyte, Ruta and Fritz, Mario}, booktitle = {International Conference on Machine Learning}, title = {Position: Safety Must Precede the Deployment of Open-Ended AI Agents}, url = {https://arxiv.org/abs/2502.04512}, year = {2026} } @misc{shi2026aevolvetraining, archiveprefix = {arXiv}, author = {Zhan Shi and Bing He and Yisi Sang and Benoit Dumoulin and Hanqing Lu}, eprint = {2606.20657}, note = {Preprint}, title = {{A-Evolve-Training: Autonomous Post-Training of a 30B Model}}, url = {https://arxiv.org/abs/2606.20657}, year = {2026} } @misc{shi2026anchor, archiveprefix = {arXiv}, author = {Dianxing Shi and Bowen Wang and Junqi He and Junhao Chen and Yuta Nakashima}, eprint = {2606.06114}, note = {Preprint}, title = {{ANCHOR: An External LLM-Driven Supervisory Module Facilitating Healthy Evolution in Self-Evolving Systems}}, url = {https://arxiv.org/abs/2606.06114}, year = {2026} } @misc{shi2026thinktwiceactprotecting, archiveprefix = {arXiv}, author = {Shanghao Shi and Xiao Wang and Chaoyu Zhang and Hao Li and Wenjing Lou and Thomas Hou and Yevgeniy Vorobeychik and Chongjie Zhang and Ning Zhang}, eprint = {2606.20922}, howpublished = {arXiv preprint arXiv:2606.20922}, primaryclass = {cs.CR}, title = {{Think Twice Before You Act: Protecting LLM Agents Against Tool Description Poisoning via Isolated Planning}}, url = {https://arxiv.org/abs/2606.20922}, year = {2026} } @misc{shinn2023reflexion, archiveprefix = {arXiv}, author = {Noah Shinn and Federico Cassano and Edward Berman and Ashwin Gopinath and Karthik Narasimhan and Shunyu Yao}, eprint = {2303.11366}, title = {{Reflexion: Language Agents with Verbal Reinforcement Learning}}, url = {https://arxiv.org/abs/2303.11366}, year = {2023} } @article{shumailov2024collapse, author = {Shumailov, Ilia and Shumaylov, Zakhar and Zhao, Yiren and Papernot, Nicolas and Anderson, Ross and Gal, Yarin}, doi = {10.1038/s41586-024-07566-y}, journal = {Nature}, pages = {755--759}, title = {{AI} Models Collapse When Trained on Recursively Generated Data}, url = {https://www.nature.com/articles/s41586-024-07566-y}, volume = {631}, year = {2024} } @misc{si2026executiongrounded, archiveprefix = {arXiv}, author = {Chenglei Si and Zitong Yang and Yejin Choi and Emmanuel Candès and Diyi Yang and Tatsunori Hashimoto}, eprint = {2601.14525}, note = {Preprint}, title = {{Towards Execution-Grounded Automated AI Research}}, url = {https://arxiv.org/abs/2601.14525}, year = {2026} } @article{silver2025welcome, author = {Silver, David and Sutton, Richard S}, journal = {Google AI}, title = {Welcome to the era of experience}, year = {2025} } @misc{singh2023restem, archiveprefix = {arXiv}, author = {Avi Singh and John D. Co-Reyes and Rishabh Agarwal and Ankesh Anand and Piyush Patil and Xavier Garcia and Peter J. Liu and James Harrison and Jaehoon Lee and Kelvin Xu and Aaron Parisi and Abhishek Kumar and Alex Alemi and Alex Rizkowsky and Azade Nova and Ben Adlam and Bernd Bohnet and Gamaleldin Elsayed and Hanie Sedghi and Igor Mordatch and Isabelle Simpson and Izzeddin Gur and Jasper Snoek and Jeffrey Pennington and Jiri Hron and Kathleen Kenealy and Kevin Swersky and Kshiteej Mahajan and Laura Culp and Lechao Xiao and Maxwell L. Bileschi and Noah Constant and Roman Novak and Rosanne Liu and Tris Warkentin and Yundi Qian and Yamini Bansal and Ethan Dyer and Behnam Neyshabur and Jascha Sohl-Dickstein and Noah Fiedel}, eprint = {2312.06585}, note = {Preprint}, title = {{Beyond Human Data: Scaling Self-Training for Problem-Solving with Language Models}}, url = {https://arxiv.org/abs/2312.06585}, year = {2023} } @misc{singh2026agentplasticity, archiveprefix = {arXiv}, author = {Harman Singh and Anton Bakhtin and Rulin Shao and Gabriel Synnaeve and Ilia Kulikov and Rob Fergus and Sanjeev Arora and Kurt Keutzer and Jason Weston and Anuj Mahajan and Anirudh Goyal}, eprint = {2610.08902}, title = {{Agent Plasticity: Measuring Self-Improvement Through Experience}}, url = {https://arxiv.org/abs/2610.08902}, year = {2026} } @misc{singh2026plasticity, archiveprefix = {arXiv}, author = {Singh, Harman and Bakhtin, Anton and Shao, Rulin and Synnaeve, Gabriel and Kulikov, Ilia and Fergus, Rob and Arora, Sanjeev and Keutzer, Kurt and Weston, Jason and Mahajan, Anuj and Goyal, Anirudh}, doi = {10.48550/arXiv.2610.08902}, eprint = {2610.08902}, primaryclass = {cs.AI}, title = {Agent Plasticity: Measuring Self-Improvement Through Experience}, url = {https://arxiv.org/abs/2610.08902}, year = {2026} } @misc{skalse2022rewardhacking, archiveprefix = {arXiv}, author = {Joar Skalse and Nikolaus H. R. Howe and Dmitrii Krasheninnikov and David Krueger}, eprint = {2209.13085}, note = {Preprint}, title = {{Defining and Characterizing Reward Hacking}}, url = {https://arxiv.org/abs/2209.13085}, year = {2022} } @inproceedings{soares2015corrigibility, author = {Soares, Nate and Fallenstein, Benja and Yudkowsky, Eliezer and Armstrong, Stuart}, booktitle = {Workshops at the Twenty-Ninth {AAAI} Conference on Artificial Intelligence}, publisher = {AAAI Publications}, title = {Corrigibility}, url = {https://intelligence.org/files/Corrigibility.pdf}, year = {2015} } @article{song2026recursivecollapse, author = {Song, Xinyuan and Cai, Zekun and Zhao, Liang}, journal = {arXiv preprint arXiv:2606.28438}, title = {When AI Reviews Its Own Code: Recursive Self-Training Collapse in Code LLMs}, url = {https://arxiv.org/abs/2606.28438}, year = {2026} } @misc{srikanth2026airesearch, archiveprefix = {arXiv}, author = {Srikanth, Dhruv and Zhao, Bingchen and Xu, Dixing and Wu, Yuxiang and Jiang, Zhengyao}, eprint = {2609.26457}, primaryclass = {cs.AI}, title = {Recursive self-improvement of AI research agents}, url = {https://arxiv.org/abs/2609.26457}, year = {2026} } @misc{srivastava2025memorygraftpersistentcompromisellm, archiveprefix = {arXiv}, author = {Saksham Sahai Srivastava and Haoyu He}, eprint = {2512.16962}, howpublished = {arXiv preprint arXiv:2512.16962}, primaryclass = {cs.CR}, title = {{MemoryGraft: Persistent Compromise of LLM Agents via Poisoned Experience Retrieval}}, url = {https://arxiv.org/abs/2512.16962}, year = {2025} } @misc{sui2026conl, archiveprefix = {arXiv}, author = {Yuan Sui and Bryan Hooi}, eprint = {2601.21464}, note = {Preprint}, title = {{Conversation for Non-verifiable Learning: Self-Evolving LLMs through Meta-Evaluation}}, url = {https://arxiv.org/abs/2601.21464}, year = {2026} } @inproceedings{sun2026equivalence, author = {Sun, Alan and Toneva, Mariya}, booktitle = {International Conference on Learning Representations}, pages = {38617--38648}, title = {{Tracking Equivalent Mechanistic Interpretations Across Neural Networks}}, url = {https://proceedings.iclr.cc/paper_files/paper/2026/hash/412e7a0143cb33d3a7a9b6abe8d90470-Abstract-Conference.html}, volume = {2026}, year = {2026} } @misc{sun2026reuse, archiveprefix = {arXiv}, author = {Xiaojing Sun and Yuhan Zeng and Zihua She and Xiao Wang}, eprint = {2609.33180}, primaryclass = {stat.ML}, title = {Which Self-Improvements Should We Trust? Reliable Self-Improvement When Agents Reuse Their Benchmarks}, url = {https://arxiv.org/abs/2609.33180}, year = {2026} } @misc{suresh2024collapserate, archiveprefix = {arXiv}, author = {Ananda Theertha Suresh and Andrew Thangaraj and Aditya Nanda Kishore Khandavally}, eprint = {2412.17646}, note = {Preprint}, title = {{Rate of Model Collapse in Recursive Training}}, url = {https://arxiv.org/abs/2412.17646}, year = {2024} } @inproceedings{tamirisa2025tamper, author = {Tamirisa, Rishub and Bharathi, Bhrugu and Phan, Long and Zhou, Andy and Gatti, Alice and Suresh, Tarun and Lin, Maxwell and Wang, Justin and Wang, Rowan and Arel, Ron and Zou, Andy and Song, Dawn and Li, Bo and Hendrycks, Dan and Mazeika, Mantas}, booktitle = {International Conference on Learning Representations}, pages = {101802--101829}, title = {{Tamper-Resistant Safeguards for Open-Weight LLMs}}, url = {https://proceedings.iclr.cc/paper_files/paper/2025/hash/fc49a629d33bc2461ed7a715ce44da68-Abstract-Conference.html}, volume = {2025}, year = {2025} } @misc{tan2026metarsirsi2, archiveprefix = {arXiv}, author = {Zihan Tan and Leixin Sun and Zitong Shi and Yitao Liu and Jiajun Wu and Nathaniel Brooks and Jiaru Qian and Xiaoran Shang and Suyuan Huang and Yi Ding and Yangxu Liao and Mukai Li and Qiushi Sun and Shudong Liu and Xuankun Rong and Xiaohang Yu and Zhuo Chen and Hejia Geng and Chenxin Li and Aozhou Wang and Zengji Tu and Robert Tang and Yuxin Zhan and Eric Jiang and Yuxin Wu and Jianqing Zhang and Xiao Liang and Fang Wu and Haochi Zhang and Alexander Marlow and Guancheng Wan}, eprint = {2609.06396}, title = {{MetaRSI / RSI2: A Meta-Recursive Self-Improving System for Recursive Self-Improving Systems Themselves}}, url = {https://arxiv.org/abs/2609.06396}, year = {2026} } @misc{tang2026fars, archiveprefix = {arXiv}, author = {Qiong Tang and Tianxiang Sun and Xiangkun Hu and Xiangyang Liu and Yiran Chen and Yunfan Shao and Bobo Li and Changze Lv and Cheng Xu and Chengsong Huang and Chunyang Li and Dizhan Xue and Hao Bai and Haodong Duan and Hengquan Guo and Hongyang He and Hongyi Chen and Hui Shen and Jiahao Yuan and Jiankai Sun and Jikang Cheng and Jinfeng Xu and Jingqi Tong and Jingye Chen and Jinxiu Liu and Jixuan Leng and Junchi Yu and Kaixun Jiang and Kun Xiang and Kunpeng Yao and Lang Feng and Liangqi Yuan and Longsen Gao and Meng Li and Qi Jia and Qiushi Sun and Shengyuan Ding and Shizhan Gong and Siru Zhong and Terry Jingchen Zhang and Tianle Gu and Tianyi Liang and Weijie Liu and Weikai Yang and Weizhi Fei and Xin Wang and Xinpeng Liu and Xuanwen Ding and Yihong Tang and Yuanli Wang and Yukun Jiang and Yuming Yang and Zhengbao He and Zhikai Chen and Zhikun Xu and Zhuang Li and Zihao Huang}, eprint = {2606.31651}, note = {Preprint}, title = {{FARS: A Fully Automated Research System Deployed at Scale}}, url = {https://arxiv.org/abs/2606.31651}, year = {2026} } @misc{tang2026generalizedagentiteratio, archiveprefix = {arXiv}, author = {Hongyao Tang and Yi Ma and Pengyi Li and Yifu Yuan}, eprint = {2609.13406}, title = {{Generalized Agent Iteration: One Formal Framework for Iterative Policy Improvement and Recursive Self-Improvement}}, url = {https://arxiv.org/abs/2609.13406}, year = {2026} } @misc{tao2024selfevolution, archiveprefix = {arXiv}, author = {Tao, Zhengwei and Lin, Ting-En and Chen, Xiancai and Li, Hangyu and Wu, Yuchuan and Li, Yongbin and Jin, Zhi and Huang, Fei and Tao, Dacheng and Zhou, Jingren}, eprint = {2404.14387}, primaryclass = {cs.CL}, title = {A Survey on Self-Evolution of Large Language Models}, url = {https://arxiv.org/abs/2404.14387}, year = {2024} } @misc{taylor2025schoolrewardhacks, archiveprefix = {arXiv}, author = {Mia Taylor and James Chua and Jan Betley and Johannes Treutlein and Owain Evans}, eprint = {2508.17511}, title = {{School of Reward Hacks: Hacking harmless tasks generalizes to misaligned behavior in LLMs}}, url = {https://arxiv.org/abs/2508.17511}, year = {2025} } @inproceedings{terekhov2026adaptivemonitors, author = {Terekhov, Mikhail and Panfilov, Alexander and Dzenhaliou, Daniil and Gulcehre, Caglar and Andriushchenko, Maksym and Prabhu, Ameya and Geiping, Jonas}, booktitle = {International Conference on Learning Representations}, pages = {51514--51565}, title = {{Adaptive Attacks on Trusted Monitors Subvert AI Control Protocols}}, url = {https://proceedings.iclr.cc/paper_files/paper/2026/hash/54b153ad8a138f4c186f21a8b7341d5e-Abstract-Conference.html}, volume = {2026}, year = {2026} } @misc{tetek2020boundedrational, archiveprefix = {arXiv}, author = {Jakub T{\v{e}}tek and Marek Sklenka and Tom{\'a}{\v{s}} Gaven{\v{c}}iak}, eprint = {2011.06275}, howpublished = {arXiv preprint arXiv:2011.06275}, title = {{Performance of Bounded-Rational Agents With the Ability to Self-Modify}}, url = {https://arxiv.org/abs/2011.06275}, year = {2020} } @misc{tian2026harnessevolutionhitsacei, archiveprefix = {arXiv}, author = {Yuan Tian and Bing Hu and Hao Wang and Binghang Lu and Fang Wu}, eprint = {2610.11655}, title = {{Harness Evolution Hits a Ceiling: When Weight Training Should Begin}}, url = {https://arxiv.org/abs/2610.11655}, year = {2026} } @inproceedings{tice2025noise, author = {Tice, Cameron and Kreer, Philipp and Helm-Burger, Nathan and Shahani, Prithviraj and Ryzhenkov, Fedor and Roger, Fabien and Neo, Clement and Haimes, Jacob and Hofst\"{a}tter, Felix and van der Weij, Teun}, booktitle = {Advances in Neural Information Processing Systems}, doi = {10.52202/085713-3922}, pages = {117590--117622}, publisher = {Curran Associates, Inc.}, title = {{Noise Injection Reveals Hidden Capabilities of Sandbagging Language Models}}, url = {https://proceedings.neurips.cc/paper_files/paper/2025/hash/aa95d031f5e151bf8fb402f44bd97054-Abstract-Conference.html}, volume = {38, Main Conference}, year = {2025} } @misc{tu2026benchguard, archiveprefix = {arXiv}, author = {Xinming Tu and Tianze Wang and Yingzhou and Lu and Kexin Huang and Yuanhao Qu and Sara Mostafavi}, eprint = {2604.24955}, note = {Preprint}, title = {{Who Guards the Benchmarks? Automated Auditing of LLM Agent Benchmarks}}, url = {https://arxiv.org/abs/2604.24955}, year = {2026} } @inproceedings{tyen2024errorlocation, author = {Gladys Tyen and Hassan Mansoor and Victor Carbune and Peter Chen and Tony Mak}, booktitle = {Findings of the Association for Computational Linguistics: ACL 2024}, doi = {10.18653/v1/2024.findings-acl.826}, pages = {13894--13908}, publisher = {Association for Computational Linguistics}, title = {{LLMs Cannot Find Reasoning Errors, but Can Correct Them Given the Error Location}}, url = {https://aclanthology.org/2024.findings-acl.826/}, year = {2024} } @inproceedings{vanderweij2025sandbagging, author = {van der Weij, Teun and Hofst\"{a}tter, Felix and Jaffe, Oliver and Brown, Samuel and Ward, Francis}, booktitle = {International Conference on Learning Representations}, pages = {73152--73189}, title = {{AI Sandbagging: Language Models can Strategically Underperform on Evaluations}}, url = {https://proceedings.iclr.cc/paper_files/paper/2025/hash/b5e5753b0a0e440a6d8dc7e143617cec-Abstract-Conference.html}, volume = {2025}, year = {2025} } @misc{wan2026deepverifier, archiveprefix = {arXiv}, author = {Yuxuan Wan and Tianqing Fang and Zaitang Li and Yintong Huo and Wenxuan Wang and Haitao Mi and Dong Yu and Michael R. Lyu}, eprint = {2601.15808}, note = {Preprint}, title = {{Inference-Time Scaling of Verification: Self-Evolving Deep Research Agents via Test-Time Rubric-Guided Verification}}, url = {https://arxiv.org/abs/2601.15808}, year = {2026} } @misc{wang2023selfconsistency, archiveprefix = {arXiv}, author = {Xuezhi Wang and Jason Wei and Dale Schuurmans and Quoc Le and Ed Chi and Sharan Narang and Aakanksha Chowdhery and Denny Zhou}, eprint = {2203.11171}, title = {{Self-Consistency Improves Chain of Thought Reasoning in Language Models}}, url = {https://arxiv.org/abs/2203.11171}, year = {2023} } @misc{wang2023voyager, archiveprefix = {arXiv}, author = {Guanzhi Wang and Yuqi Xie and Yunfan Jiang and Ajay Mandlekar and Chaowei Xiao and Yuke Zhu and Linxi Fan and Anima Anandkumar}, eprint = {2305.16291}, title = {{Voyager: An Open-Ended Embodied Agent with Large Language Models}}, url = {https://arxiv.org/abs/2305.16291}, year = {2023} } @misc{wang2025agenticeval, archiveprefix = {arXiv}, author = {Yixu Wang and Xin Wang and Yang Yao and Xinyuan Li and Xibang Yang and Yan Teng and Xingjun Ma and Yingchun Wang}, eprint = {2509.26100}, howpublished = {arXiv preprint arXiv:2509.26100}, title = {{AgenticEval: Toward Agentic and Self-Evolving Safety Evaluation of Large Language Models}}, url = {https://arxiv.org/abs/2509.26100}, year = {2025} } @inproceedings{wang2025agentmemoryprivacy, address = {Vienna, Austria}, author = {Wang, Bo and He, Weiyi and Zeng, Shenglai and Xiang, Zhen and Xing, Yue and Tang, Jiliang and He, Pengfei}, booktitle = {Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers)}, doi = {10.18653/v1/2025.acl-long.1227}, month = {July}, pages = {25241--25260}, publisher = {Association for Computational Linguistics}, title = {Unveiling Privacy Risks in {LLM} Agent Memory}, url = {https://aclanthology.org/2025.acl-long.1227/}, year = {2025} } @inproceedings{wang2025fairlycertain, author = {Wang, Yinong Oliver and Sivakumar, Nivedha and Arif Khan, Falaah and Metcalf, Katherine and Golinski, Adam and Mackraz, Natalie and Theobald, Barry-John and Zappella, Luca and Apostoloff, Nicholas}, booktitle = {Proceedings of the 42nd International Conference on Machine Learning}, pages = {64343--64366}, publisher = {PMLR}, series = {Proceedings of Machine Learning Research}, title = {{Is Your Model Fairly Certain? Uncertainty-Aware Fairness Evaluation for LLMs}}, url = {https://proceedings.mlr.press/v267/wang25cp.html}, volume = {267}, year = {2025} } @misc{wang2025unveilingprivacyrisksllm, archiveprefix = {arXiv}, author = {Bo Wang and Weiyi He and Shenglai Zeng and Zhen Xiang and Yue Xing and Jiliang Tang and Pengfei He}, eprint = {2502.13172}, howpublished = {arXiv preprint arXiv:2502.13172}, primaryclass = {cs.CR}, title = {{Unveiling Privacy Risks in LLM Agent Memory}}, url = {https://arxiv.org/abs/2502.13172}, year = {2025} } @misc{wang2026auditingharnesstamperingselfimproving, archiveprefix = {arXiv}, author = {Xing Wang and Xiaoyi Zhang and Jie Shao}, eprint = {2609.00069}, howpublished = {arXiv preprint arXiv:2609.00069}, primaryclass = {cs.CL}, title = {{Auditing Harness Tampering in Self-Improving Agents}}, url = {https://arxiv.org/abs/2609.00069}, year = {2026} } @misc{wang2026benchmarkaudit, archiveprefix = {arXiv}, author = {Junlin Wang and Federico Bianchi and Shang Zhu and Fan Nie and Yongchan Kwon and Bhuwan Dhingra and James Zou}, eprint = {2605.26079}, note = {Preprint}, title = {{Automated Benchmark Auditing for AI Agents and Large Language Models}}, url = {https://arxiv.org/abs/2605.26079}, year = {2026} } @article{wang2026diving, author = {Wang, Kaiqi and Hou, Wenjin and Yan, Yuchen and Jia, Hongrui and Zhong, Zhisheng and Ren, Botao and Chen, Yifei and Zhang, Songyang and Shen, Yongliang and Xiao, Jun and Yang, Yi and Zhuang, Yueting and Fan, Hehe}, journal = {OpenReview.net}, title = {Diving into Reliable Self-Evolving Agents: A Survey}, url = {https://openreview.net/forum?id=CGO1hDTHNe}, year = {2026} } @misc{wang2026dynamicrubric, archiveprefix = {arXiv}, author = {Beining Wang and Weihang Su and Hongtao Tian and Hao Kong and Tao Yang and Ting Yao and Qingyi Pan and Yueyue Wu and Qingyao Ai and Min Zhang and Yiqun Liu}, eprint = {2607.20083}, note = {Preprint}, title = {{Co-Evolving LLM Evaluators and Policies via DynamicRubric}}, url = {https://arxiv.org/abs/2607.20083}, year = {2026} } @misc{wang2026evohunt, archiveprefix = {arXiv}, author = {Ziyue Wang and Cheuk Wang Maurice Ng and Chenchen Yu and Strick Sheng and Kaihua Qin and Liyi Zhou}, eprint = {2606.16420}, note = {Preprint}, title = {{Transferable Self-Evolving Playbooks for Agentic Security Auditing}}, url = {https://arxiv.org/abs/2606.16420}, year = {2026} } @misc{wang2026harnessevaluation, archiveprefix = {arXiv}, author = {Yike Wang and Huaisheng Zhu and Zhengyu Hu and Yige Yuan and Zhengyu Chen and Shakti Senthil and Hannaneh Hajishirzi and Yulia Tsvetkov and Pradeep Dasigi and Teng Xiao}, eprint = {2607.12227}, title = {{Rethinking the Evaluation of Harness Evolution for Agents}}, url = {https://arxiv.org/abs/2607.12227}, year = {2026} } @inproceedings{wang2026huxley, archiveprefix = {arXiv}, author = {Wenyi Wang and Piotr Pi{\k{e}}kos and Li Nanbo and Firas Laakom and Yimeng Chen and Mateusz Ostaszewski and Mingchen Zhuge and J{\"u}rgen Schmidhuber}, booktitle = {International Conference on Learning Representations}, eprint = {2510.21614}, title = {{Huxley-G{\"o}del Machine: Human-Level Coding Agent Development by an Approximation of the Optimal Self-Improving Machine}}, url = {https://arxiv.org/abs/2510.21614}, year = {2026} } @misc{wang2026mempotdefendingmemoryextraction, archiveprefix = {arXiv}, author = {Yuhao Wang and Shengfang Zhai and Guanghao Jin and Yinpeng Dong and Linyi Yang and Jiaheng Zhang}, eprint = {2602.07517}, howpublished = {arXiv preprint arXiv:2602.07517}, primaryclass = {cs.CR}, title = {{MemPot: Defending Against Memory Extraction Attack with Optimized Honeypots}}, url = {https://arxiv.org/abs/2602.07517}, year = {2026} } @misc{wang2026oeppoisoningselfevolvingllm, archiveprefix = {arXiv}, author = {Kaixiang Wang and Jiong Lou and Zhaojiacheng Zhou and Jie Li}, eprint = {2605.18930}, howpublished = {arXiv preprint arXiv:2605.18930}, primaryclass = {cs.CR}, title = {{OEP: Poisoning Self-Evolving LLM Agents via Locally Correct but Non-Transferable Experiences}}, url = {https://arxiv.org/abs/2605.18930}, year = {2026} } @misc{wang2026rememberrevealprivacyawarememory, archiveprefix = {arXiv}, author = {Wenjie Wang and Wenhe Si and Xinyue Xu and Yue Xu}, eprint = {2608.16551}, howpublished = {arXiv preprint arXiv:2608.16551}, primaryclass = {cs.CR}, title = {{What to Remember, What to Reveal: Privacy-Aware Memory for Conversational Agents}}, url = {https://arxiv.org/abs/2608.16551}, year = {2026} } @misc{wang2026selfconsuming, archiveprefix = {arXiv}, author = {Yaxuan Wang and Zhongteng Cai and Yujia Bao and Xueru Zhang and Yang Liu}, eprint = {2601.05184}, howpublished = {arXiv preprint arXiv:2601.05184}, title = {{Observations and Remedies for Large Language Model Bias in Self-Consuming Performative Loop}}, url = {https://arxiv.org/abs/2601.05184}, year = {2026} } @misc{wang2026serpo, archiveprefix = {arXiv}, author = {Jianze Wang and Kunwang Zheng and Ying Liu and Yu Cao and Qilong Zhang and Jinlong Chen and Hua Yang and Qianglong Chen}, eprint = {2607.26873}, note = {Preprint}, title = {{SERPO: Self-Evolving Rubric Policy Optimization for Open-Ended Test-Time Reinforcement Learning}}, url = {https://arxiv.org/abs/2607.26873}, year = {2026} } @misc{wang2026whenbettergetsworse, archiveprefix = {arXiv}, author = {Ke Wang and Zijie Zhao and Zhiyi Yuan and Changlun Li}, eprint = {2609.32677}, title = {{When Better Gets Worse: Improvement Fidelity for Self-Improving Agents in Adaptive Worlds}}, url = {https://arxiv.org/abs/2609.32677}, year = {2026} } @inproceedings{wen2025adaptivecontrol, author = {Wen, Jiaxin and Hebbar, Vivek and Larson, Caleb and Bhatt, Aryan and Radhakrishnan, Ansh and Sharma, Mrinank and Sleight, Henry and Feng, Shi and He, He and Perez, Ethan and Shlegeris, Buck and Khan, Akbir}, booktitle = {International Conference on Learning Representations}, pages = {42650--42693}, title = {{Adaptive Deployment of Untrusted LLMs Reduces Distributed Threats}}, url = {https://proceedings.iclr.cc/paper_files/paper/2025/hash/699c5a2bb87a8aa32c4ad541d9997361-Abstract-Conference.html}, volume = {2025}, year = {2025} } @misc{wen2026contextualizedprivacydefensellm, archiveprefix = {arXiv}, author = {Yule Wen and Yanzhe Zhang and Jianxun Lian and Xiaoyuan Yi and Xing Xie and Diyi Yang}, eprint = {2603.02983}, howpublished = {arXiv preprint arXiv:2603.02983}, primaryclass = {cs.CR}, title = {{Contextualized Privacy Defense for LLM Agents}}, url = {https://arxiv.org/abs/2603.02983}, year = {2026} } @misc{wijk2024rebench, archiveprefix = {arXiv}, author = {Wijk, Hjalmar and Lin, Tao and Becker, Joel and Jawhar, Sami and Parikh, Neev and Broadley, Thomas and Chan, Lawrence and Chen, Michael and Clymer, Josh and Dhyani, Jai and Ericheva, Elena and Garcia, Katharyn and Goodrich, Brian and Jurkovic, Nikola and Karnofsky, Holden and Kinniment, Megan and Lajko, Aron and Nix, Seraphina and Sato, Lucas and Saunders, William and Taran, Maksym and West, Ben and Barnes, Elizabeth}, eprint = {2411.15114}, title = {{RE-Bench: Evaluating frontier AI R\&D capabilities of language model agents against human experts}}, url = {https://arxiv.org/abs/2411.15114}, year = {2024} } @inproceedings{williams2025manipulation, author = {Williams, Marcus and Carroll, Micah and Narang, Adhyyan and Weisser, Constantin and Murphy, Brendan and Dragan, Anca}, booktitle = {International Conference on Learning Representations}, pages = {89542--89593}, title = {{On Targeted Manipulation and Deception when Optimizing LLMs for User Feedback}}, url = {https://proceedings.iclr.cc/paper_files/paper/2025/hash/deed907f507e526a8dd20896d2bd7c49-Abstract-Conference.html}, volume = {2025}, year = {2025} } @misc{wu2024metarewarding, archiveprefix = {arXiv}, author = {Tianhao Wu and Weizhe Yuan and Olga Golovneva and Jing Xu and Yuandong Tian and Jiantao Jiao and Jason Weston and Sainbayar Sukhbaatar}, eprint = {2407.19594}, note = {Preprint}, title = {{Meta-Rewarding Language Models: Self-Improving Alignment with LLM-as-a-Meta-Judge}}, url = {https://arxiv.org/abs/2407.19594}, year = {2024} } @misc{wu2024reversal, archiveprefix = {arXiv}, author = {Ting Wu and Xuefeng Li and Pengfei Liu}, eprint = {2407.05013}, note = {Preprint}, title = {{Progress or Regress? Self-Improvement Reversal in Post-training}}, url = {https://arxiv.org/abs/2407.05013}, year = {2024} } @inproceedings{wu2025axbench, author = {Wu, Zhengxuan and Arora, Aryaman and Geiger, Atticus and Wang, Zheng and Huang, Jing and Jurafsky, Dan and Manning, Christopher D and Potts, Christopher}, booktitle = {Proceedings of the 42nd International Conference on Machine Learning}, pages = {67035--67080}, publisher = {PMLR}, title = {{{A}x{B}ench: Steering {LLM}s? {E}ven Simple Baselines Outperform Sparse Autoencoders}}, url = {https://proceedings.mlr.press/v267/wu25a.html}, volume = {267}, year = {2025} } @inproceedings{wu2025unlearned, author = {Wu, Xiaoyu and Pang, Yifei and Liu, Terrance and Wu, Steven}, booktitle = {Advances in Neural Information Processing Systems}, doi = {10.52202/085713-1226}, editor = {D. Belgrave and C. Zhang and H. Lin and R. Pascanu and P. Koniusz and M. Ghassemi and N. Chen}, pages = {36524--36553}, publisher = {Curran Associates, Inc.}, title = {{Unlearned but Not Forgotten: Data Extraction after Exact Unlearning in LLM}}, url = {https://proceedings.neurips.cc/paper_files/paper/2025/file/33ffb8c041f43e592793fbfef90ff720-Paper-Conference.pdf}, volume = {38, Main Conference}, year = {2025} } @article{wu2026ai4ai, author = {Wu, Kai and Lyu, Hao and Luo, Zhen and Wang, Chaofan and Ye, Siyu and Lin, Jinghao and Ji, Xiaozhong and Jiang, Boyuan and Wang, Shengzhi and Wang, Zihan and Ye, Yiwen and Wang, Hao and Wang, Zimu and Liu, Wenzhe and Wang, Ruobing and Cai, Kai and Xiong, Mingliang and Fang, Wen and Liu, Mingqing and Zhang, Yifan and Yang, Lei and Hu, Xiaobin and Liu, Qingwen}, doi = {10.20944/preprints202608.2108.v1}, journal = {Preprints.org}, title = {{AI4AI} Survey: From Long-Horizon Agents to Recursive Self-Improvement---Definitions, Reliable Horizons, and Open Problems}, url = {https://www.preprints.org/manuscript/202608.2108}, year = {2026} } @inproceedings{wu2026evolver, archiveprefix = {arXiv}, author = {Rong Wu and Xiaoman Wang and Jianbiao Mei and Pinlong Cai and Daocheng Fu and Cheng Yang and Licheng Wen and Xuemeng Yang and Yufan Shen and Yuxin Wang and Botian Shi}, booktitle = {International Conference on Machine Learning}, eprint = {2510.16079}, title = {{EvolveR: Self-Evolving LLM Agents through an Experience-Driven Lifecycle}}, url = {https://arxiv.org/abs/2510.16079}, year = {2026} } @misc{wu2026evomalselfpoisoningselfevolvingcoding, archiveprefix = {arXiv}, author = {Xiaodong Wu and Yu Shi and Qi Li and Zhimin Zhao and Xiangman Li and Bram Adams and Ahmed E. Hassan and Jianbing Ni}, eprint = {2608.25776}, howpublished = {arXiv preprint arXiv:2608.25776}, primaryclass = {cs.CR}, title = {{EVOMAL: Self-Poisoning in Self-Evolving Coding Agents}}, url = {https://arxiv.org/abs/2608.25776}, year = {2026} } @inproceedings{wu2026interruptible, author = {Wu, Tsung-Han and Miroyan, Mihran and Chan, David M. and Darrell, Trevor and Norouzi, Narges and Gonzalez, Joseph E.}, booktitle = {Proceedings of the 43rd International Conference on Machine Learning}, pages = {136539--136581}, publisher = {PMLR}, title = {{Are Large Reasoning Models Interruptible?}}, url = {https://proceedings.mlr.press/v306/wu26ao.html}, volume = {306}, year = {2026} } @misc{wu2026medrsirecursiveselfimprovementmedical, archiveprefix = {arXiv}, author = {Junde Wu and Jiayuan Zhu and Minghao Hu and Fenglin Liu and Jiazhen Pan}, eprint = {2609.24838}, primaryclass = {cs.AI}, title = {MedRSI: Recursive Self-Improvement for Medical Agents via Clinically Aligned Self-Evolution}, url = {https://arxiv.org/abs/2609.24838}, year = {2026} } @misc{wu2026modularrsi, archiveprefix = {arXiv}, author = {Wu, Siwei and Ren, Jincheng and Li, Yizhi and Li, Haau-Sing and Yang, Chengran and Zhang, Yuxuan and Gu, Weicheng and Yang, Jian and Batista-Navarro, Riza and Zhang, Chuanyi and Liu, Xianglong and Zhou, Ming and Dai, Bryan and Lin, Chenghua}, eprint = {2609.14857}, primaryclass = {cs.CL}, title = {{ModularRSI}: Modular and Generalizable Recursive Harness Self-Improvement}, url = {https://arxiv.org/abs/2609.14857}, year = {2026} } @misc{wu2026sgm, archiveprefix = {arXiv}, author = {Xuening Wu and Shenqin Yin and Yanlan Kang and Xinhang Zhang and Qianya Xu and Zeping Chen and Wenqiang Zhang}, eprint = {2510.10232}, title = {{SGM: A Statistical Godel Machine for Risk-Controlled Recursive Self-Modification}}, url = {https://arxiv.org/abs/2510.10232}, year = {2026} } @inproceedings{wu2026socialbias, author = {Addison J. Wu and Ryan Liu and Xuechunzi Bai and Thomas L. Griffiths}, booktitle = {International Conference on Machine Learning}, title = {Large Language Models Develop Novel Social Biases Through Adaptive Exploration}, url = {https://arxiv.org/abs/2511.06148}, year = {2026} } @misc{wu2026stv, archiveprefix = {arXiv}, author = {Chen Henry Wu and Aditi Raghunathan}, eprint = {2605.30290}, note = {Preprint}, title = {{Self-Trained Verification for Training- and Test-Time Self-Improvement}}, url = {https://arxiv.org/abs/2605.30290}, year = {2026} } @misc{wyllie2024fairnessloops, archiveprefix = {arXiv}, author = {Sierra Wyllie and Ilia Shumailov and Nicolas Papernot}, eprint = {2403.07857}, note = {Preprint}, title = {{Fairness Feedback Loops: Training on Synthetic Data Amplifies Bias}}, url = {https://arxiv.org/abs/2403.07857}, year = {2024} } @misc{xia2026rrsi, archiveprefix = {arXiv}, author = {Xia, Peng and Han, Rujun and Wang, Zifeng and Chen, Yanfei and Zhuang, Yufan and Lee, Yoonho and Huang, Chengsong and Yu, Han and CuiZhu, Zhongying and Ming, Yifei and Yao, Huaxiu and Gokturk, Burak and Pfister, Tomas and Lee, Chen-Yu}, eprint = {2609.24972}, primaryclass = {cs.LG}, title = {{RRSI}: Regularized Recursive Self-Improvement of Agent Harnesses}, url = {https://arxiv.org/abs/2609.24972}, year = {2026} } @inproceedings{xiang2025guardagent, author = {Xiang, Zhen and Zheng, Linzhi and Li, Yanjie and Hong, Junyuan and Li, Qinbin and Xie, Han and Zhang, Jiawei and Xiong, Zidi and Xie, Chulin and Yang, Carl and Song, Dawn and Li, Bo}, booktitle = {Proceedings of the 42nd International Conference on Machine Learning}, pages = {68316--68342}, publisher = {PMLR}, title = {{{G}uard{A}gent: Safeguard {LLM} Agents via Knowledge-Enabled Reasoning}}, url = {https://proceedings.mlr.press/v267/xiang25a.html}, volume = {267}, year = {2025} } @misc{xiang2026agentslearnyoubenchmarking, archiveprefix = {arXiv}, author = {Yongli Xiang and Zhifang Zhang and Bojun Yang and Ziming Hong and Lei Feng and Miao Xu and Tongliang Liu}, eprint = {2608.03700}, howpublished = {arXiv preprint arXiv:2608.03700}, primaryclass = {cs.CR}, title = {{When Agents Learn to Be You: Benchmarking Privacy Leakage, Impersonation Risk, and Defenses in Persona Skills}}, url = {https://arxiv.org/abs/2608.03700}, year = {2026} } @inproceedings{xiao2026air, archiveprefix = {arXiv}, author = {Zibo Xiao and Jun Sun and Junjie Chen}, booktitle = {International Conference on Machine Learning}, eprint = {2602.11749}, title = {{AIR: Improving Agent Safety through Incident Response}}, url = {https://arxiv.org/abs/2602.11749}, year = {2026} } @inproceedings{xie2025bone, author = {Xie, Guofu and Zhang, Xiao and Yao, Ting and Shi, Yunsheng}, booktitle = {Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics}, pages = {27237--27263}, title = {Bone soups: {A} seek-and-soup model merging approach for controllable multi-objective generation}, year = {2025} } @misc{xie2026canagentsforgetwhenasked, archiveprefix = {arXiv}, author = {Bofan Xie and Chaokun Wang and Guanchen Ge and Songyao Wang and Fang Niu and Hang Zhang and Wei Wang}, eprint = {2610.18052}, title = {{Can Agents Forget When Asked? A Benchmark for Memory Unlearning in LLM Agents}}, url = {https://arxiv.org/abs/2610.18052}, year = {2026} } @inproceedings{xiong2025draftfaithfulness, author = {Xiong, Zidi and Chen, Shan and Qi, Zhenting and Lakkaraju, Himabindu}, booktitle = {Advances in Neural Information Processing Systems}, doi = {10.52202/085713-4655}, pages = {139438--139467}, publisher = {Curran Associates, Inc.}, title = {{Measuring the Faithfulness of Thinking Drafts in Large Reasoning Models}}, url = {https://proceedings.neurips.cc/paper_files/paper/2025/hash/cbd25a91123291348cc8407a38b75080-Abstract-Conference.html}, volume = {38, Main Conference}, year = {2025} } @misc{xiong2026alma, archiveprefix = {arXiv}, author = {Yiming Xiong and Shengran Hu and Jeff Clune}, eprint = {2602.07755}, title = {{Learning to Continually Learn via Meta-learning Agentic Memory Designs}}, url = {https://arxiv.org/abs/2602.07755}, year = {2026} } @inproceedings{xiong2026freegift, author = {Xiong, Zidi and Chen, Shan and Lakkaraju, Himabindu}, booktitle = {Proceedings of the 43rd International Conference on Machine Learning}, pages = {140231--140252}, publisher = {PMLR}, title = {{Monitorability as a Free Gift: How {RLVR} Spontaneously Aligns Reasoning}}, url = {https://proceedings.mlr.press/v306/xiong26b.html}, volume = {306}, year = {2026} } @misc{xu2026asievolve, archiveprefix = {arXiv}, author = {Weixian Xu and Tiantian Mi and Yixiu Liu and Yang Nan and Zhimeng Zhou and Lyumanshan Ye and Lin Zhang and Yu Qiao and Pengfei Liu}, eprint = {2603.29640}, note = {Preprint}, title = {{ASI-Evolve: AI Accelerates AI}}, url = {https://arxiv.org/abs/2603.29640}, year = {2026} } @misc{xu2026drafting, archiveprefix = {arXiv}, author = {Xu, Xin and Cheng, Ke and Wang, Hao and Huang, Yangru and Xiao, Yanghao and Chen, Zhichao and Zhong, Zhisheng and Liu, Xun and Li, Yi and Li, Qian and Lv, Haibin and Zhang, Jun and Jiang, Jie and Lin, Zhouchen}, doi = {10.48550/arXiv.2610.16491}, eprint = {2610.16491}, primaryclass = {cs.LG}, title = {Drafting Agents, Not Just Tokens: Speculative Recursive Self-Improvement for Agentic Systems}, url = {https://arxiv.org/abs/2610.16491}, year = {2026} } @misc{xu2026draftingagentsnotjusttok, archiveprefix = {arXiv}, author = {Xin Xu and Ke Cheng and Hao Wang and Yangru Huang and Yanghao Xiao and Zhichao Chen and Zhisheng Zhong and Xun Liu and Yi Li and Qian Li and Haibin Lv and Jun Zhang and Jie Jiang and Zhouchen Lin}, eprint = {2610.16491}, title = {{Drafting Agents, Not Just Tokens: Speculative Recursive Self-Improvement for Agentic Systems}}, url = {https://arxiv.org/abs/2610.16491}, year = {2026} } @misc{xu2026evo, archiveprefix = {arXiv}, author = {Chejian Xu and Shujian Zhang and Wenxuan Zhou and John Lambert and Beck Chen and Bo Li and Andrew Hard and Rajiv Mathews}, eprint = {2610.13025}, title = {{Evo$^2$: Evolving the Evolver through Weight-Harness Co-Evolution}}, url = {https://arxiv.org/abs/2610.13025}, year = {2026} } @misc{xu2026evo2, archiveprefix = {arXiv}, author = {Xu, Chejian and Zhang, Shujian and Zhou, Wenxuan and Lambert, John and Chen, Beck and Li, Bo and Hard, Andrew and Mathews, Rajiv}, doi = {10.48550/arXiv.2610.13025}, eprint = {2610.13025}, primaryclass = {cs.CL}, title = {{Evo$^2$}: Evolving the Evolver through Weight-Harness Co-Evolution}, url = {https://arxiv.org/abs/2610.13025}, year = {2026} } @misc{xu2026rise, archiveprefix = {arXiv}, author = {Chaoran Xu and Yingmao Miao and Pengfei Zhang and Hao Dou and Lei Sun and Xiangxiang Chu}, eprint = {2605.20914}, note = {Preprint}, title = {{RISE: Reliable Improvement in Self-Evolving Vision-Language Models}}, url = {https://arxiv.org/abs/2605.20914}, year = {2026} } @misc{xu2026unlearningisntdeletioninvestigating, archiveprefix = {arXiv}, author = {Xiaoyu Xu and Xiang Yue and Yang Liu and Qingqing Ye and Huadi Zheng and Peizhao Hu and Minxin Du and Haibo Hu}, eprint = {2505.16831}, howpublished = {arXiv preprint arXiv:2505.16831}, primaryclass = {cs.CL}, title = {{Unlearning Isn't Deletion: Investigating Reversibility of Machine Unlearning in LLMs}}, url = {https://arxiv.org/abs/2505.16831}, year = {2026} } @misc{xue2026pastbench, archiveprefix = {arXiv}, author = {Shuhan Xue and Zixin Ding and Yichen Shen and Yinjie Wang and Zhenfei Yin and Yingcheng Wu and Yuxin Chen and Mengdi Wang and Ling Yang}, eprint = {2608.04003}, note = {Preprint}, title = {{PAST-Bench: Benchmarking the Foundations of Recursive Self-Improvement in Personal Agents}}, url = {https://arxiv.org/abs/2608.04003}, year = {2026} } @misc{xue2026sciencebuddy, archiveprefix = {arXiv}, author = {Xue, Shuhan and Zhong, Jianyuan and Nan, Ziyuan and Li, Wenbin and Yu, Zhaochen and Ding, Jinchao and Gao, Qiang and Zhan, Pengyu and Zhang, Yuntong and Cheng, Tian and Yin, Zhenfei and Wu, Yingcheng and Yang, Ling}, eprint = {2609.17523}, primaryclass = {cs.AI}, title = {{ScienceBuddy}: Recursive-in-Recursive Self-Improvement for Interactive Scientific Agents}, url = {https://arxiv.org/abs/2609.17523}, year = {2026} } @misc{yamada2025aiscientistv2, archiveprefix = {arXiv}, author = {Yutaro Yamada and Robert Tjarko Lange and Cong Lu and Shengran Hu and Chris Lu and Jakob Foerster and Jeff Clune and David Ha}, eprint = {2504.08066}, note = {Preprint}, title = {{The AI Scientist-v2: Workshop-Level Automated Scientific Discovery via Agentic Tree Search}}, url = {https://arxiv.org/abs/2504.08066}, year = {2025} } @misc{yan2026benignaloneharmfultogeth, archiveprefix = {arXiv}, author = {Bingyu Yan and Xiaoming Zhang and Chaozhuo Li and Ziyi Zhou and Yirui Qi and Litian Zhang}, eprint = {2608.01759}, title = {{Benign Alone, Harmful Together: Exploiting Experience Composition in Self-Evolving LLM Agents}}, url = {https://arxiv.org/abs/2608.01759}, year = {2026} } @misc{yan2026pacevolve, archiveprefix = {arXiv}, author = {Yan, Minghao and Peng, Bo and Coleman, Benjamin and Chen, Ziqi and Xie, Zhouhang and Chen, Shuo and He, Zhankui and Sachdeva, Noveen and Ye, Isabella and Wang, Weili and Wang, Chi and Chi, Ed H. and Pereira, Fernando and Kang, Wang-Cheng and Cheng, Derek Zhiyuan and Wang, Beidou}, eprint = {2601.10657}, note = {Preprint}, title = {{PACEvolve}: Enabling Progress-Aware Consistent Evolution}, url = {https://arxiv.org/abs/2601.10657}, year = {2026} } @inproceedings{yang2025superficial, author = {Yang, Wenkai and Shen, Shiqi and Shen, Guangyao and Yao, Wei and Liu, Yong and Zhi, Gong and Lin, Yankai and Wen, Ji-Rong}, booktitle = {International Conference on Learning Representations}, pages = {3271--3295}, title = {{Super(ficial)-alignment: Strong Models May Deceive Weak Models in Weak-to-Strong Generalization}}, url = {https://proceedings.iclr.cc/paper_files/paper/2025/hash/092359ce5cf60a80e882378944bf1be4-Abstract-Conference.html}, volume = {2025}, year = {2025} } @misc{yang2026frontis, archiveprefix = {arXiv}, author = {Junlin Yang and Che Jiang and Yu Fu and Tianwei Luo and Can Ren and Weizhi Wang and Kaikai Zhao and Hongyi Liu and Yuxin Zuo and Yuru Wang and Yuchen Fan and Kai Tian and Zhenzhao Yuan and Xiaojian Lin and Li Sheng and Rushi Qiang and Guoli Jia and Xingtai Lv and Ermo Hua and Dianqiao Lei and Youbang Sun and Ning Ding and Bowen Zhou and Kaiyan Zhang}, eprint = {2607.28568}, title = {{Frontis-MA1: Training an AI4AI Model towards Recursive Self-Improvement in Machine Learning Engineering}}, url = {https://arxiv.org/abs/2607.28568}, year = {2026} } @misc{yang2026skillopt, archiveprefix = {arXiv}, author = {Yifan Yang and Ziyang Gong and Weiquan Huang and Qihao Yang and Ziwei Zhou and Zisu Huang and Yan Li and Xuemei Gao and Qi Dai and Bei Liu and Kai Qiu and Yuqing Yang and Dongdong Chen and Xue Yang and Chong Luo}, eprint = {2605.23904}, note = {Preprint}, title = {{SkillOpt: Executive Strategy for Self-Evolving Agent Skills}}, url = {https://arxiv.org/abs/2605.23904}, year = {2026} } @misc{yang2026spurious, archiveprefix = {arXiv}, author = {Yiwei Yang and Haoxiang Zhang and Bingbing Wen and Yao Lu and Yuchen Wu and Lei Zhang and Julian McAuley and Pan Lu and Bill Howe}, eprint = {2609.16268}, note = {Preprint}, primaryclass = {cs.CL}, title = {Spurious Tool Use: When {RL} Agents Learn the Wrong Reason to Act}, url = {https://arxiv.org/abs/2609.16268}, year = {2026} } @misc{yang2026vaso, archiveprefix = {arXiv}, author = {Yunhao Yang and Neel P. Bhatt and Kevin Wang and Samuel Tetteh and Zhangyang Wang and Ufuk Topcu}, eprint = {2606.05395}, title = {{VASO: Formally Verifiable Self-Evolving Skills for Physical AI Agents}}, url = {https://arxiv.org/abs/2606.05395}, year = {2026} } @misc{yang2026zombie, archiveprefix = {arXiv}, author = {Xianglin Yang and Yufei He and Shuo Ji and Bryan Hooi and Jin Song Dong}, eprint = {2602.15654}, howpublished = {arXiv preprint arXiv:2602.15654}, primaryclass = {cs.CR}, title = {{Zombie Agents: Persistent Control of Self-Evolving LLM Agents via Self-Reinforcing Injections}}, url = {https://arxiv.org/abs/2602.15654}, year = {2026} } @misc{ye2026fragility, archiveprefix = {arXiv}, author = {Ye, Qinyuan and Li, Yu and Pruksachatkun, Yada and Zhang, Jiaxin and Wu, Chien-Sheng}, eprint = {2608.18066}, note = {Preprint}, title = {On the Fragility of Self-Improving Agents: Variance, Task Order, and Underspecification}, url = {https://arxiv.org/abs/2608.18066}, year = {2026} } @misc{yi2025escapingcollapse, archiveprefix = {arXiv}, author = {Bingji Yi and Qiyuan Liu and Yuwei Cheng and Haifeng Xu}, eprint = {2510.16657}, note = {Preprint}, title = {{Escaping Model Collapse via Synthetic Data Verification: Near-term Improvements and Long-term Convergence}}, url = {https://arxiv.org/abs/2510.16657}, year = {2025} } @misc{yin2023selfknowledge, archiveprefix = {arXiv}, author = {Zhangyue Yin and Qiushi Sun and Qipeng Guo and Jiawen Wu and Xipeng Qiu and Xuanjing Huang}, eprint = {2305.18153}, note = {Preprint}, title = {{Do Large Language Models Know What They Don't Know?}}, url = {https://arxiv.org/abs/2305.18153}, year = {2023} } @inproceedings{yin2025godel, author = {Xunjian Yin and Xinyi Wang and Liangming Pan and Li Lin and Xiaojun Wan and William Yang Wang}, booktitle = {Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers)}, doi = {10.18653/v1/2025.acl-long.1354}, pages = {27890--27913}, publisher = {Association for Computational Linguistics}, title = {{G{\"o}del Agent: A Self-Referential Agent Framework for Recursively Self-Improvement}}, url = {https://aclanthology.org/2025.acl-long.1354/}, year = {2025} } @misc{yin2026onpolicy, archiveprefix = {arXiv}, author = {Bo Yin and Qi Li and Xinchao Wang}, eprint = {2605.11882}, howpublished = {arXiv preprint arXiv:2605.11882}, title = {{On-Policy Self-Evolution via Failure Trajectories for Agentic Safety Alignment}}, url = {https://arxiv.org/abs/2605.11882}, year = {2026} } @inproceedings{yin2026partitioned, author = {Yin, Ren and Ishida, Takashi and Sugiyama, Masashi}, booktitle = {International Conference on Learning Representations}, pages = {112525--112548}, title = {{Towards Scalable Oversight via Partitioned Human Supervision}}, url = {https://proceedings.iclr.cc/paper_files/paper/2026/hash/b6f03ce953cf7f07dafa20891d2539e3-Abstract-Conference.html}, volume = {2026}, year = {2026} } @misc{ying2026evolvingdeception, archiveprefix = {arXiv}, author = {Zonghao Ying and Haowen Dai and Tianyuan Zhang and Yisong Xiao and Quanchen Zou and Aishan Liu and Jian Yang and Yaodong Yang and Xianglong Liu}, eprint = {2603.05872}, title = {{Evolving Deception: When Agents Evolve, Deception Wins}}, url = {https://arxiv.org/abs/2603.05872}, year = {2026} } @misc{ying2026skilljackpersistentskillbackdoors, archiveprefix = {arXiv}, author = {Zonghao Ying and Xiangfan Wu and Huiyu Wu and Xing Zheng and Huangsheng Cheng and Xiaorong Shi and Jing Guo}, eprint = {2608.03509}, howpublished = {arXiv preprint arXiv:2608.03509}, primaryclass = {cs.CR}, title = {{SkillJack: Persistent Skill Backdoors in Self-Evolving Agents}}, url = {https://arxiv.org/abs/2608.03509}, year = {2026} } @inproceedings{yu2025trustworthyagents, author = {Yu, Miao and Meng, Fanci and Zhou, Xinyun and Wang, Shilong and Mao, Junyuan and Pan, Linsey and Chen, Tianlong and Wang, Kun and Li, Xinfeng and Zhang, Yongfeng and An, Bo and Wen, Qingsong}, booktitle = {Proceedings of the 31st ACM SIGKDD Conference on Knowledge Discovery and Data Mining}, doi = {10.1145/3711896.3736561}, pages = {6216--6226}, publisher = {Association for Computing Machinery}, title = {A Survey on Trustworthy LLM Agents: Threats and Countermeasures}, year = {2025} } @misc{yu2026forget, archiveprefix = {arXiv}, author = {Yu, Ye and Yuan, Xiaopeng and Jin, Haibo and Liu, Heming and Yu, Yaoning and Wang, Haohan}, eprint = {2605.09315}, primaryclass = {cs.AI}, title = {Do Self-Evolving Agents Forget? Capability Degradation and Preservation in Lifelong LLM Agent Adaptation}, url = {https://arxiv.org/abs/2605.09315}, year = {2026} } @misc{yu2026harmonyharness, archiveprefix = {arXiv}, author = {Weichen Yu and Lesly Miculicich and Bhavana Dalvi Mishra and Jiefeng Chen and Shuo Li and Matt Fredrikson and Tomas Pfister and Long T. Le}, eprint = {2610.15150}, title = {{Harmony Harness: Multi-Objective Self-Evolution of Agent Harnesses}}, url = {https://arxiv.org/abs/2610.15150}, year = {2026} } @misc{yu2026latentskill, archiveprefix = {arXiv}, author = {Aofan Yu and Chenyu Zhou and Tianyi Xu and Zihan Guo and Rong Shan and Zhihui Fu and Jun Wang and Weiwen Liu and Yong Yu and Weinan Zhang and Jianghao Lin}, eprint = {2606.06087}, title = {{LatentSkill: From In-Context Textual Skills to In-Weight Latent Skills for LLM Agents}}, url = {https://arxiv.org/abs/2606.06087}, year = {2026} } @misc{yuan2025selfrewarding, archiveprefix = {arXiv}, author = {Weizhe Yuan and Richard Yuanzhe Pang and Kyunghyun Cho and Xian Li and Sainbayar Sukhbaatar and Jing Xu and Jason Weston}, eprint = {2401.10020}, title = {{Self-Rewarding Language Models}}, url = {https://arxiv.org/abs/2401.10020}, year = {2025} } @misc{yuan2026seva, archiveprefix = {arXiv}, author = {Aojie Yuan and Yi Nian and Haiyue Zhang and Zijian Su and Yue Zhao}, eprint = {2606.29713}, note = {Preprint}, title = {{SEVA: Self-Evolving Verification Agent with Process Reward for Fact Attribution}}, url = {https://arxiv.org/abs/2606.29713}, year = {2026} } @misc{yudkowsky2013tiling, author = {Eliezer Yudkowsky and Marcello Herreshoff}, howpublished = {Machine Intelligence Research Institute}, note = {Early draft}, title = {{Tiling Agents for Self-Modifying AI, and the L{\"o}bian Obstacle}}, url = {https://intelligence.org/files/TilingAgents.pdf}, year = {2013} } @inproceedings{yue2025reasoningcapacity, author = {Yue, Yang and Chen, Zhiqi and Lu, Rui and Zhao, Andrew and Wang, Zhaokai and Yue, Yang and Song, Shiji and Huang, Gao}, booktitle = {Advances in Neural Information Processing Systems}, doi = {10.52202/085713-1933}, pages = {57654--57689}, publisher = {Curran Associates, Inc.}, title = {Does Reinforcement Learning Really Incentivize Reasoning Capacity in {LLMs} Beyond the Base Model?}, url = {https://papers.nips.cc/paper_files/paper/2025/hash/537d5aa768c2d534016a4d06f87bc8fb-Abstract-Conference.html}, volume = {38, Main Conference}, year = {2025} } @inproceedings{zelikman2024stop, author = {Zelikman, Eric and Lorch, Eliana and Mackey, Lester and Kalai, Adam Tauman}, booktitle = {Conference on Language Modeling}, title = {Self-Taught Optimizer (STOP): Recursively Self-Improving Code Generation}, url = {https://openreview.net/forum?id=46Zgqo4QIU}, year = {2024} } @inproceedings{zhang2024patching, author = {Zhang, Fred and Nanda, Neel}, booktitle = {International Conference on Learning Representations}, pages = {1651--1678}, title = {{Towards Best Practices of Activation Patching in Language Models: Metrics and Methods}}, url = {https://proceedings.iclr.cc/paper_files/paper/2024/hash/06a52a54c8ee03cd86771136bc91eb1f-Abstract-Conference.html}, volume = {2024}, year = {2024} } @misc{zhang2025ace, archiveprefix = {arXiv}, author = {Qizheng Zhang and Changran Hu and Shubhangi Upasani and Boyuan Ma and Fenglu Hong and Vamsidhar Kamanuru and Jay Rainton and Chen Wu and Mengmeng Ji and Hanchen Li and Urmish Thakker and James Zou and Kunle Olukotun}, eprint = {2510.04618}, note = {Preprint}, title = {{Agentic Context Engineering: Evolving Contexts for Self-Improving Language Models}}, url = {https://arxiv.org/abs/2510.04618}, year = {2025} } @inproceedings{zhang2025control, author = {Zhang, Yi-Fan and Zhang, Xiao and Zhang, Min-Ling}, booktitle = {Proceedings of the 42nd International Conference on Machine Learning}, pages = {77017--77043}, title = {Generalization analysis for controllable learning}, year = {2025} } @inproceedings{zhang2025controllable, author = {Jingyu Zhang and Ahmed Elgohary Ghoneim and Ahmed Magooda and Daniel Khashabi and Ben Van Durme}, booktitle = {International Conference on Learning Representations}, title = {{Controllable Safety Alignment: Inference-Time Adaptation to Diverse Safety Requirements}}, url = {https://proceedings.iclr.cc/paper_files/paper/2025/hash/a9fa03d8b6b0564580337c985ad10a04-Abstract-Conference.html}, year = {2025} } @inproceedings{zhang2025dgm, archiveprefix = {arXiv}, author = {Jenny Zhang and Shengran Hu and Cong Lu and Robert Lange and Jeff Clune}, booktitle = {International Conference on Learning Representations}, eprint = {2505.22954}, title = {{Darwin G{\"o}del Machine: Open-Ended Evolution of Self-Improving Agents}}, url = {https://arxiv.org/abs/2505.22954}, year = {2026} } @misc{zhang2025memevolve, archiveprefix = {arXiv}, author = {Guibin Zhang and Haotian Ren and Chong Zhan and Zhenhong Zhou and Junhao Wang and He Zhu and Wangchunshu Zhou and Shuicheng Yan}, eprint = {2512.18746}, title = {{MemEvolve: Meta-Evolution of Agent Memory Systems}}, url = {https://arxiv.org/abs/2512.18746}, year = {2025} } @inproceedings{zhang2025whowhen, archiveprefix = {arXiv}, author = {Zhang, Shaokun and Yin, Ming and Zhang, Jieyu and Liu, Jiale and Han, Zhiguang and Zhang, Jingyang and Li, Beibin and Wang, Chi and Wang, Huazheng and Chen, Yiran and Wu, Qingyun}, booktitle = {Proceedings of the 42nd International Conference on Machine Learning}, eprint = {2505.00212}, pages = {76583--76599}, publisher = {PMLR}, title = {{Which Agent Causes Task Failures and When? {O}n Automated Failure Attribution of {LLM} Multi-Agent Systems}}, url = {https://proceedings.mlr.press/v267/zhang25cq.html}, volume = {267}, year = {2025} } @inproceedings{zhang2026agentracer, author = {Zhang, Guibin and Wang, Junhao and Chen, Junjie and Zhou, Wangchunshu and Wang, Kun and YAN, Shuicheng}, booktitle = {International Conference on Learning Representations}, pages = {11377--11399}, title = {{AgenTracer: Who Is Inducing Failure in the LLM Agentic Systems?}}, url = {https://proceedings.iclr.cc/paper_files/paper/2026/hash/134ed7a477770f227f12450ef0cbb8f4-Abstract-Conference.html}, volume = {2026}, year = {2026} } @misc{zhang2026coevoskills, archiveprefix = {arXiv}, author = {Hanrong Zhang and Shicheng Fan and Henry Peng Zou and Yankai Chen and Zhenting Wang and Jiayu Zhou and Chengze Li and Wei-Chieh Huang and Yifei Yao and Kening Zheng and Xue and Liu and Xiaoxiao Li and Philip S. Yu}, eprint = {2604.01687}, note = {Preprint}, title = {{CoEvoSkills: Self-Evolving Agent Skills via Co-Evolutionary Verification}}, url = {https://arxiv.org/abs/2604.01687}, year = {2026} } @misc{zhang2026credit, archiveprefix = {arXiv}, author = {Haiyue Zhang}, eprint = {2608.19760}, note = {Preprint}, title = {{Credit Without Ground Truth: Auditing Step-Level Credit Assignment in LLM Agents Against Executed Replay}}, url = {https://arxiv.org/abs/2608.19760}, year = {2026} } @misc{zhang2026grader, archiveprefix = {arXiv}, author = {Xing Zhang and Guanghui Wang and Yanwei Cui and Ziyuan Li and Wei Qiu and Bing Zhu and Peiyang He}, eprint = {2607.12790}, title = {{Who Grades the Grader? Co-Evolving Evaluation Metrics and Skills for Self-Improving LLM Agents}}, url = {https://arxiv.org/abs/2607.12790}, year = {2026} } @misc{zhang2026harnesscompass, archiveprefix = {arXiv}, author = {Luan Zhang and Ruochen Zhou and Dandan Song and Zhengyu Chen and Yuhang Tian and Jun Yang and Huipeng Ma and Chenhao Li and Guangyuan Feng and Xudong Li and Yizhou Jin and Yan Xu}, eprint = {2608.01918}, title = {{HarnessCompass: Guiding Automatic Harness Evolution toward Generalizable and Effective Agent Harnesses}}, url = {https://arxiv.org/abs/2608.01918}, year = {2026} } @misc{zhang2026hyperagents, archiveprefix = {arXiv}, author = {Jenny Zhang and Bingchen Zhao and Wannan Yang and Jakob Foerster and Jeff Clune and Minqi Jiang and Sam Devlin and Tatiana Shavrina}, eprint = {2603.19461}, title = {{Hyperagents}}, url = {https://arxiv.org/abs/2603.19461}, year = {2026} } @inproceedings{zhang2026interventionvalue, archiveprefix = {arXiv}, author = {Chubin Zhang and Zhenglin Wan and Xingrui Yu and jingxuan wu and Qi Wen and Pengfei Zhou and Wangbo Zhao and Ivor Tsang}, booktitle = {Advances in Neural Information Processing Systems}, eprint = {2606.21399}, note = {Accepted at NeurIPS 2026; proceedings forthcoming}, title = {{Calibration Is Not Control: Intervention Advantage for LLM-Agent Oversight}}, url = {https://neurips.cc/virtual/2026/poster/149357}, year = {2026} } @article{zhang2026librarydrift, author = {Zhang, Xing and Cui, Yanwei and Wang, Guanghui and Li, Ziyuan and Qiu, Wei and Zhu, Bing and He, Peiyang}, journal = {arXiv preprint arXiv:2605.19576}, title = {Library Drift: Diagnosing and Fixing a Silent Failure Mode in Self-Evolving LLM Skill Libraries}, url = {https://arxiv.org/abs/2605.19576}, year = {2026} } @misc{zhang2026privacypeekauditingllmbasedagents, archiveprefix = {arXiv}, author = {Mingxuan Zhang and Jiahui Han and Dadi Guo and Songze Li and Guanchu Wang and Na Zou and Dongrui Liu and Xia Hu}, eprint = {2606.00152}, howpublished = {arXiv preprint arXiv:2606.00152}, primaryclass = {cs.CR}, title = {{PrivacyPeek: Auditing What LLM-Based Agents Acquire, Not Just What They Say}}, url = {https://arxiv.org/abs/2606.00152}, year = {2026} } @misc{zhang2026safetysurviveselfimprovementfailures, archiveprefix = {arXiv}, author = {Yunbei Zhang and Janet Wang and Saiyue Lyu and Yingqiang Ge and Kaiqu Liang and Zijian Jin and Chandan K Reddy and Jihun Hamm}, eprint = {2610.01073}, primaryclass = {cs.SE}, title = {Safety Must Survive Self-Improvement: Why Failures Persist and How Agents Recover}, url = {https://arxiv.org/abs/2610.01073}, year = {2026} } @misc{zhang2026selaur, archiveprefix = {arXiv}, author = {Dengjia Zhang and Xiaoou Liu and Lu Cheng and Yaqing Wang and Kenton Murray and Hua Wei}, eprint = {2602.21158}, note = {Preprint}, title = {{SELAUR: Self Evolving LLM Agent via Uncertainty-aware Rewards}}, url = {https://arxiv.org/abs/2602.21158}, year = {2026} } @misc{zhang2026selfharness, archiveprefix = {arXiv}, author = {Hangfan Zhang and Shao Zhang and Kangcong Li and Chen Zhang and Yang Chen and Yiqun Zhang and Lei Bai and Shuyue Hu}, eprint = {2606.09498}, note = {Preprint}, title = {{Self-Harness: Harnesses That Improve Themselves}}, url = {https://arxiv.org/abs/2606.09498}, year = {2026} } @misc{zhang2026sigma, archiveprefix = {arXiv}, author = {Jingyu Zhang and Shruti Palaskar and Daniel Khashabi and Benjamin Van Durme and Leon A. Gatys and Joseph Yitan Cheng}, eprint = {2610.07935}, title = {{SIGMA: Self-Improving Alignment Generalization from a Model Spec}}, url = {https://arxiv.org/abs/2610.07935}, year = {2026} } @misc{zhang2026skilltolora, archiveprefix = {arXiv}, author = {Zhang, Tianyi and Qi, Zhonghao}, eprint = {2606.16769}, title = {{Skill-to-LoRA: From Using Skills to Learning Behaviors for Token-Efficient LLM Agents}}, url = {https://arxiv.org/abs/2606.16769}, year = {2026} } @misc{zhang2026spf, archiveprefix = {arXiv}, author = {Jiawen Zhang and Yangfan Hu and Kejia Chen and Lipeng He and Jiachen Ma and Jian Lou and Dan Li and Jian Liu and Xiaohu Yang and Ruoxi Jia}, eprint = {2601.10141}, title = {{Understanding and Preserving Safety in Fine-Tuned LLMs}}, url = {https://arxiv.org/abs/2601.10141}, year = {2026} } @misc{zhang2026tempo, archiveprefix = {arXiv}, author = {Qingyang Zhang and Xinke Kong and Haitao Wu and Qinghua Hu and Minghao Wu and Baosong Yang and Yu Cheng and Yun Luo and Ganqu Cui and Changqing Zhang}, eprint = {2604.19295}, note = {Preprint}, title = {{TEMPO: Scaling Test-time Training for Large Reasoning Models}}, url = {https://arxiv.org/abs/2604.19295}, year = {2026} } @misc{zhao2024expel, archiveprefix = {arXiv}, author = {Andrew Zhao and Daniel Huang and Quentin Xu and Matthieu Lin and Yong-Jin Liu and Gao Huang}, eprint = {2308.10144}, title = {{ExpeL: LLM Agents Are Experiential Learners}}, url = {https://arxiv.org/abs/2308.10144}, year = {2024} } @inproceedings{zhao2026faithful, author = {Weixiang Zhao and Yingshuo Wang and Yichen Zhang and Yang Deng and Yanyan Zhao and Wanxiang Che and Bing Qin and Ting Liu}, booktitle = {International Conference on Machine Learning}, title = {Large Language Model Agents Are Not Always Faithful Self-Evolvers}, url = {https://arxiv.org/abs/2601.22436}, year = {2026} } @misc{zhao2026parametric, archiveprefix = {arXiv}, author = {Xuan Zhao and Haonan He and Qingyu Yang and Minglei Li and Jingqi Ye and Zelin Tan and Bo Wan and Peng Ye}, eprint = {2606.30015}, title = {{Parametric Skills}}, url = {https://arxiv.org/abs/2606.30015}, year = {2026} } @inproceedings{zhao2026safetyrisks, address = {San Diego, California, United States}, author = {Zhao, Weixiang and Zhang, Yichen and Wang, Yingshuo and Deng, Yang and Zhao, Yanyan and Zhi, Xuda and Huang, Yongbo and He, Hao and Che, Wanxiang and Qin, Bing and Liu, Ting}, booktitle = {Findings of the Association for Computational Linguistics: ACL 2026}, doi = {10.18653/v1/2026.findings-acl.2091}, month = {July}, pages = {42145--42169}, publisher = {Association for Computational Linguistics}, title = {On Safety Risks in Experience-Driven Self-Evolving Agents}, url = {https://aclanthology.org/2026.findings-acl.2091/}, year = {2026} } @misc{zheng2026benchshieldformalmodelbackedinstrumentation, archiveprefix = {arXiv}, author = {Shenghan Zheng and Zonglin Di and Yimin Liu and Kyoung Whan Choe and Jiankai Sun and Heguang Lin and Penghao Jiang and Yifeng He and Xiao Cheng and Jicheng Wang and Wenbo Chen and Alex Yates and Yinzhe Zhao and Bingran You and Yuan Gao and Ayush Munot and Shubham Gaur and Zhe Ye and Hao Wang and Xiangyi Li and Dawn Song and Christophe Hauser}, eprint = {2609.11028}, howpublished = {arXiv preprint arXiv:2609.11028}, primaryclass = {cs.CR}, title = {{BenchShield: Formal Model-Backed Instrumentation for Reward Integrity in LLM-Agent Evaluation Infrastructure}}, url = {https://arxiv.org/abs/2609.11028}, year = {2026} } @article{zheng2026dreamrsi, author = {Zheng, Tong and Wu, Xidong and Zhang, Zheng and He, Zhankui and Zhang, Chaoyi and Coleman, Benjamin and Wei, Ruoqiao and Bai, Di and Liu, Haolin and Liu, Rui and Wang, Xue and Zhuan, Yue and Kang, Wang-Cheng and Xiang, Renkai and Huang, Heng and Cheng, Xinwu and Guo, Yunsong}, journal = {arXiv preprint arXiv:2609.14858}, title = {Dream-RSI: Recursive Self-Improvement through Evolving Worlds}, url = {https://arxiv.org/abs/2609.14858}, year = {2026} } @misc{zheng2026seagym, archiveprefix = {arXiv}, author = {Congjie Zheng and Chuanyi Xue and Bin Liang and Jun Yang and Changshui Zhang}, eprint = {2606.17546}, note = {Preprint}, title = {{SEAGym: An Evaluation Environment for Self-Evolving LLM Agents}}, url = {https://arxiv.org/abs/2606.17546}, year = {2026} } @inproceedings{zhong2026weightwatch, author = {Zhong, Ziqian and Raghunathan, Aditi}, booktitle = {International Conference on Learning Representations}, pages = {75198--75234}, title = {{Watch the Weights: Unsupervised monitoring and control of fine-tuned LLMs}}, url = {https://proceedings.iclr.cc/paper_files/paper/2026/hash/79ee64e62fd3ee763961ce3d6aed4971-Abstract-Conference.html}, volume = {2026}, year = {2026} } @misc{zhou2024lats, archiveprefix = {arXiv}, author = {Andy Zhou and Kai Yan and Michal Shlapentokh-Rothman and Haohan Wang and Yu-Xiong Wang}, eprint = {2310.04406}, title = {{Language Agent Tree Search Unifies Reasoning Acting and Planning in Language Models}}, url = {https://arxiv.org/abs/2310.04406}, year = {2024} } @misc{zhou2026athanor, archiveprefix = {arXiv}, author = {Yufa Zhou and Yichen Wang and Jiakun Fan and Ziqi Shu and Zhichao Liu and Anru R. Zhang and Neiwen Ling and Zijun Wei}, eprint = {2610.21055}, title = {{Athanor: An Agent-Native System for Autonomous Post-Training Research}}, url = {https://arxiv.org/abs/2610.21055}, year = {2026} } @misc{zhou2026selfevolvingcoding, archiveprefix = {arXiv}, author = {Zhou, Hao and Hu, Haichuan and Luo, Tianyu and Shang, Ye and Fang, Chunrong and Chen, Zhenyu and Xiao, Liang and Zhang, Quanjun}, eprint = {2608.03392}, primaryclass = {cs.SE}, title = {Self-Evolving Coding Agents}, url = {https://arxiv.org/abs/2608.03392}, year = {2026} } @misc{zhou2026traceaudit, archiveprefix = {arXiv}, author = {Xiaolin Zhou and Jinbo Liu and Li Li and Ryan A. Rossi and Xiyang Hu}, eprint = {2605.11946}, title = {{Counterfactual Trace Auditing of LLM Agent Skills}}, url = {https://arxiv.org/abs/2605.11946}, year = {2026} } @misc{zhou2026verifine, archiveprefix = {arXiv}, author = {Zewei Zhou and Rachel Luo and Yulong Cao and Chaowei Xiao and Chensheng Peng and Boyi Li and Thomas Tian and Zheng Lian and Yan Wang and Jiaqi Ma and Boris Ivanovic and Marco Pavone and Wenhao Ding}, eprint = {2610.08761}, title = {{VeriFine: Scaling Verification for Self-Improvement in Embodied Reasoning}}, url = {https://arxiv.org/abs/2610.08761}, year = {2026} } @misc{zhu2025agentdebug, archiveprefix = {arXiv}, author = {Kunlun Zhu and Zijia Liu and Bingxuan Li and Muxin Tian and Yingxuan Yang and Jiaxun Zhang and Pengrui Han and Qipeng Xie and Fuyang Cui and Weijia Zhang and Xiaoteng Ma and Xiaodong Yu and Gowtham Ramesh and Jialian Wu and Zicheng Liu and Pan Lu and James Zou and Jiaxuan You}, eprint = {2509.25370}, title = {{Where LLM Agents Fail and How They can Learn From Failures}}, url = {https://arxiv.org/abs/2509.25370}, year = {2025} } @misc{zhu2026aed, archiveprefix = {arXiv}, author = {Zhu, Kunlun and Ye, Xuyan and Li, Yibo and Qian, Cheng and Li, Beibin and Ji, Heng}, eprint = {2609.40111}, title = {{Agent Error Dataset: Scaling 50,000 Error--Diagnosis Pairs for Failure Analysis and Error-Aware Post-Training}}, url = {https://arxiv.org/abs/2609.40111}, year = {2026} } @misc{zhu2026rsiagent, archiveprefix = {arXiv}, author = {Sibo Zhu and Shicheng Fan and Xinyue Wang and Wenyi Wu and Kun Zhou and Biwei Huang}, eprint = {2609.15364}, note = {Preprint}, title = {{RSIAgent: Autonomous Exploration for Recursive Self-improvement in New Environments}}, url = {https://arxiv.org/abs/2609.15364}, year = {2026} } @misc{zhu2026skillcoach, archiveprefix = {arXiv}, author = {Jiayin Zhu and Kelong Mao and Yudong Guo and Dengbo He and Sulong Xu and Simiu Gu and Yutao Yue}, eprint = {2607.01874}, note = {Preprint}, title = {{SkillCoach: Self-Evolving Rubrics for Evaluating and Enhancing Agentic Skill-Use}}, url = {https://arxiv.org/abs/2607.01874}, year = {2026} } @misc{zhu2026stealthapartharmtogether, archiveprefix = {arXiv}, author = {Zihao Zhu and Siwei Lyu and Adel Bibi and Baoyuan Wu}, eprint = {2609.30383}, howpublished = {arXiv preprint arXiv:2609.30383}, primaryclass = {cs.AI}, title = {{Stealth Apart, Harm Together: Skill Cascading Attacks on Skill-Based Agent Systems}}, url = {https://arxiv.org/abs/2609.30383}, year = {2026} } @inproceedings{zweiger2025seal, author = {Adam Zweiger and Jyo Pari and Han Guo and Yoon Kim and Pulkit Agrawal}, booktitle = {Advances in Neural Information Processing Systems}, doi = {10.52202/085713-2483}, title = {Self-Adapting Language Models}, url = {https://papers.neurips.cc/paper_files/paper/2025/hash/6b41e04c41726e2a60e456d0a2b961ab-Abstract-Conference.html}, volume = {38}, year = {2025} }