Improvement Validity
Whether reported gains are genuine, retained, and transferable under valid evaluation.
RESEARCH SURVEY / TRUSTWORTHY AI
Trustworthy Recursive Self-Improvementin the Era of Large Models
Foundations, Risks, Evaluation, Applications, and Open Challenges
Assessing improvement validity, constraint preservation, and effective oversight across successive updates, successor systems, and recursive revisions to improvement mechanisms.
4Trustworthiness themes
3Levels of trust claims
8Survey chapters
263Classified records
EXPLORE THE SURVEYFour entry points. One connected framework.
01 / FOUNDATIONS ↗
RSI Process
Step through S1–S4 and recursive revision.
Explore the process →02 / TRUSTWORTHINESS ↗Risks & Safeguards
Four themes linked to the chapter reading lists.
Explore the themes →03 / EVALUATION ↗Evidence & Claims
Compare update, successor, and meta-level scope.
Compare claim levels →04 / LITERATURE ↗Literature Explorer
Search, save, share, and export references.
Browse 413 records →01 / SURVEY OVERVIEW
Trustworthiness under
Recursive Self-Improvement
Self-modification alone does not guarantee trustworthy improvement. Gains, constraints, and oversight must be reassessed as systems and their improvement mechanisms change.
Systems built around large models increasingly use experience and feedback to make persistent changes to models, prompts, memories, tools, and workflows. Recursive self-improvement (RSI) extends this process to inductive biases and to the mechanisms that govern future updates.
This survey presents an operational definition and process framework for RSI, and examines risks, safeguards, and evaluation evidence for individual updates, effects inherited by successor systems, and revisions to improvement mechanisms. It distinguishes direct RSI findings from evidence in related systems.
Reading guide · Manuscript §1 and §2.3 · Analytical roles describe what is assessed; they are not stages of the improvement process.
Constraint Preservation
Whether behavioral and information protections remain effective as changes are inherited and reused.
Effective Oversight
Whether explanations, verification, and authorized intervention remain useful after recursive revision.
02 / PRELIMINARIES
The RSI Process
S1–S4 revise the policy h or inductive bias B. Recursive improvement (RI) applies the same stages to the bias update function U itself.
Manuscript §2.1 · Policy h, inductive bias B, and bias update function U have distinct roles. Improvement is an objective, not a guaranteed result.
OPERATIONAL DEFINITION
RSI is a process in which a learning system not only updates its policy using new experience but also autonomously updates its inductive bias. The mechanism that updates this bias is itself recursively revised, with the aim of sustained gains in task performance.
Policyht → ht+1
Inductive biasBt → Bt+1
Bias update functionUt → Ut+1
Accepted revisions are retained for subsequent cycles.03 / ORGANIZING FRAMEWORK
Four Trustworthiness Themes
The themes specify what must be assessed or protected. S1–S4 and RI locate where requirements may be violated and where safeguards may be applied.
04 / RISKS, SAFEGUARDS, AND EVIDENCE
Trustworthiness Requirements
Four interconnected themes under three complementary analytical roles.
JUMP TO A READING LIST
CHAPTER 3Improvement Validity
Genuine and
Sustained Improvement
Validity, reliability, robustness, and truthfulness. Distinguish genuine gains from evaluation artifacts, and assess retention and transfer across subsequent versions.
CHAPTER 4Constraint Preservation
Behavioral Constraints,
Alignment, and Fairness
Behavioral safety, value alignment, and group fairness. Examine whether adaptations preserve relevant human goals and constraints across inherited changes.
CHAPTER 5Constraint Preservation
System Security, Resilience,
Privacy, and Data Governance
Protect system integrity and sensitive information. Examine poisoned artifacts, privacy leakage, and the persistence of compromised state in successor systems.
CHAPTER 6Effective Oversight
Transparency
and Controllability
Transparency, interpretability, human control, accountability, and governance. Reassess explanations, correction, and oversight after changes to the system and its updater.
05 / EVALUATING TRUSTWORTHY RSI
Three Levels of
Trustworthiness Claims
Update, successor, and meta-improvement claims concern different objects and require different evidence. They are not additional stages of the RSI process.
WHAT THE EVIDENCE DOES NOT ESTABLISH
A successful update does not establish trustworthy recursive improvement.
Evidence must name the update, lineage, or updater tested, together with its task scope, assessment horizon, and untested conditions. At the meta-improvement level, both the quality of later updates and the validity of their checks and controls require reassessment.
Examine the open problems →UPDATE LEVEL
The current change is effective.
Compare an accepted revision with the unchanged system under comparable conditions. Use evidence independent of candidate development and check the integrated result against applicable requirements.
Evaluation requirementsIndependent assessment
Matched conditions
Constraint checks
Matched conditions
Constraint checks
SUCCESSOR LEVEL
The properties survive inheritance.
Assess frozen successor snapshots longitudinally. Test newly targeted tasks alongside previously acquired capabilities, behavioral constraints, information protections, and oversight.
Evaluation requirementsFrozen snapshots
Retention and transfer
Longitudinal assessment
Retention and transfer
Longitudinal assessment
META-IMPROVEMENT LEVEL
The revised updater produces better subsequent updates.
Compare Ut and Ut+1 through paired continuations from a common policy, bias, and persistent state. Match task streams, permissions, and total budgets; assess later updates, constraints, and oversight independently.
Evaluation requirementsPaired continuations
Fixed-updater controls
Renewed verification
Fixed-updater controls
Renewed verification
06 / LITERATURE AND BIBLIOGRAPHY
Literature Explorer
Download all BibTeX ↓Search 413 bibliography records, including 263 classified records across eight chapters. Save a reading list, share your filters, or export the current results.
/
07 / APPLICATIONS AND OPEN PROBLEMS
Open Problems
and Future Directions
Application context changes the evidence required: what the system can modify, who controls evaluation feedback, and what persists after acceptance.
What can the system change?
Persistent artifacts, system behavior and implementation, or the improvement mechanism itself.
Who controls the test?
Self-authored evaluation, co-evolving evaluators, or external and protected evaluation.
What happens after acceptance?
Reversible software changes, persistent derived state, or consequential real-world actions.
Manuscript §8.3 · The seven directions below preserve the manuscript’s headings; descriptions are condensed for the website.
01Theoretical Foundations of Trustworthiness under Recursive Change+
Develop accounts of how trustworthiness requirements change when the policy, inductive bias, and bias update function evolve together.
02Verifying Self-Modifying Improvement Mechanisms+
Assess whether verification assumptions, acceptance rules, and authorized controls remain valid when the mechanisms producing future updates are revised.
03Long-Horizon Risk Dynamics and Adaptive Mitigation+
Study how risks persist, interact, or change across successor systems, and evaluate mitigation over longer improvement trajectories.
04Recursive Improvement of Trustworthiness Mechanisms+
Examine revisions to evaluators, monitors, and safeguards themselves, with independent evidence of their effects on subsequent updates.
05From Test-Time Scaling to Test-Time Controlling+
Investigate how runtime control and intervention can accompany additional test-time computation and adaptation.
06Fundamental Trade-offs and Limits of Trustworthy RSI+
Characterize the limits and trade-offs among improvement, constraint preservation, verification cost, and effective oversight.
07Domain-Specific and Multi-Stakeholder Trustworthiness+
Adapt evaluation and governance to the requirements of different domains, affected populations, and stakeholders.
SURVEY RESOURCESExplore the Literature.
Explore the Literature.
Contribute to the Survey.
Browse the chapter-level taxonomy, reference metadata, and source bibliography.