acceptodds
Under review as a conference paper at ICLR 2027

Learning to Repair the Learner: Evaluating Recursive Improvement in Neuro-Symbolic World Models

Abstract

Self-improving agents are usually credited through descendant-performance curves: later lineage members outscore their ancestors. Such curves misattribute improvement even in a fully inspectable system. We formalize a checkpoint separating world model, repair procedure, evidence and a fixed kernel, and derive a ladder of matched interventions, from crossed model/procedure states and owned versus replayed evidence to a code×history crossing, a generator lock and non-recursive baselines. We apply it to a neuro-symbolic lineage that edits its repair procedure and successor generator, on an original cohort and replications with 252 fresh families declared before execution. In replication, inherited procedure explains 45% of the naive descendant gain; model structure explains the rest. Under a shared source history, continued editing detectably improves on freezing, but through final code (history effect ≈ 0), and is practically equivalent to choosing one fixed behavior per family on the same continuation evidence with fewer fits; generator editing is practically equivalent to locking it. A positive control shows the ladder detects planted generator and history value. Because the base generator indexes proposals by history length, sharing a history between source lineages decides whether checkpoints lock into single-term repair (504/504 versus 75/504); with separate histories the gain over freezing is no longer detected. A replicated gain thus hinges on the checkpoint’s history contract, the failure our formalism exists to expose.

Primary area:
transfer learning, meta learning, and lifelong learning
Keywords:
recursive self-improvement, world models, neuro-symbolic AI, counterfactual credit assignment, meta learning, self-modifying agents, world model repair, matched interventions, causal attribution, non-recursive baselines, generator plasticity, checkpoint semantics, history dependence, equivalence testing, model-based planning

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_4ZLGOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper examines how to attribute apparent recursive self-improvement in agents that repair executable world models. It introduces a checkpoint formalism and an evidence ladder comprising model/procedure crossings, owned versus replayed evidence, code/history crossings, generator locks, and non-recursive baselines. Experiments in an inspectable synthetic system, including fresh replications with 252 world families, show that descendant gains combine inherited model structure and procedure changes. Continued editing is practically equivalent to an evidence-matched non-recursive baseline, and its advantage over freezing depends strongly on the source-history contract.

Critical strengths
  1. The paper clearly distinguishes descendant quality from improver quality. The checkpoint-based interventions make this distinction concrete, while the history-contract experiment shows how frozen-versus-continued comparisons can be misleading.
  2. The experiments are transparent and carefully controlled. Replications, paired uncertainty estimates, equivalence tests, influence analyses, and an evidence-matched non-recursive baseline strengthen the findings.
Critical weaknesses
  1. The attribution framework is evaluated only in the authors’ own testbed. The paper does not apply its attribution measures to established settings such as STOP or DGM. Evaluating those settings with the proposed framework would show whether it can reveal attribution patterns beyond the task and environment designed by the authors [1, 2].
  2. The conclusion about generator editing is not fully supported by the main evaluation. In the main test, the generator is not executed during final evaluation, so the experiment cannot directly show whether editing it improves later descendants. The authors should evaluate generated descendants directly before concluding that generator editing has little value .
  3. The benefit of continued editing depends on how history is shared. The benefit appears with shared histories but disappears with separate histories, because history length changes which edits are proposed. The authors should test additional history settings and clarify which one best represents the intended system before generalizing this result .

[1] Zelikman, E., Lorch, E., Mackey, L., and Kalai, A. T. "Self-Taught Optimizer (STOP): Recursively Self-Improving Code Generation." Conference on Language Modeling, 2024.

[2] Zhang, J., Hu, S., Lu, C., Lange, R., and Clune, J. "Darwin Godel Machine: Open-Ended Evolution of Self-Improving Agents." International Conference on Learning Representations, 2026.

Questions for authors

How does the proposed evidence ladder provide additional insight beyond existing self-improvement evaluations such as STOP and DGM? In particular, which conclusion in this paper could not be obtained from descendant-performance comparisons or frozen-improver controls alone?

1 comment
AuthorsOpenReview ↗

Separating code and history effects and testing the generator on descendants

Thank you for pressing us to say what a reader can learn beyond the controls in STOP and DGM. Your observation that the generator does not run during frozen transfer also helped us clarify why the two generator evaluations support different conclusions.

Figure 3 (p. 7) gives a concrete answer to your first question. Continued editing beat freezing by 0.01010 mean loss on shared source histories. Either the resulting code or the records carried to transfer could explain that difference. Running each final program with each final history gave a code contrast of −0.01029 and a history contrast of +0.00019, with exact history ties in 247/252 families. A comparison of the two unmodified endpoints cannot recover those contrasts. STOP already compares seed and evolved improvers from the same starting solution on a downstream task, and tests transfer to held-out tasks. DGM includes a fixed-modifier ablation. Section 2 credits those controls. A code-by-history crossing on either platform would be a new external experiment.

Section 5.3 of the original submission included a direct test of descendants. Each final program performs two editing rounds using the transfer task’s train/validation split before protected scoring. The natural continued-minus-generator-locked effect was −0.00260 (95% interval [−0.00622, +0.00044]). Its 90% interval crosses the −0.005 equivalence boundary. The revised abstract and Table 4 (p. 8) therefore describe natural generator benefit under live editing as inconclusive. The frozen-transfer equivalence result concerns what earlier edits leave in the final repairer and history.

Separate writable histories represent autonomous learners. A shared history represents a team permitted to read one another’s trials (§6, p. 9). We added an exploratory scheduling intervention on the same 126 families. When the base proposer advanced by stage, the interaction between the two history contracts fell from +0.00782 to +0.00092 (§5.4 and Appendix E.7). This tests the proposal rule under both ownership contracts. Validation on STOP, DGM or another external system remains open.

Reviewer_gDpsOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary

The main goal of the paper is to investigate how "self-improving" AI agents are evaluated, arguing that standard descendant-performance curves (showing newer AI models beating older ones) can be misleading because they often credit wrong factors.

Core contributions:

  • The paper establishes a formal checkpointing system that separates the world model, repair procedures, history, and evaluation kernels.
  • It introduces an "evidence ladder" of controlled interventions (such as crossed model/procedure states, code history crossings, and generator locks) in order to rigorously test self-improvement claims.
  • It provides an inspectable neuro-symbolic testbed that the authors have designed to run hundreds of independent family tests quickly.
Critical strengths

The paper presents a very rigorous methodological framework (its evidence ladder):

  • This is consequential because upsofar, prior work evaluating self-improving AI agents typically relied on simplistic descendant-performance curves, which could easily mask whether an agent was actually improving itself or simply benefiting from inherited model structures and pre-supplied evidence. Because the paper formalizes a checkpoint structure and introduces a ladder of matched interventions, the authors provide a rigorous "diagnostic blueprint". This can fundamentally shift the standard of proof for recursive self-improvement claims, and make future research much more accountable.

The paper explores empirical falsifications well:

  • Rather than presenting a single isolated success story, the authors tested their claims across many robust cohorts, including pre-declared, pre-registered replications with hundreds of fresh families (n=252n=252, as they state). Their controlled experiments successfully overturned initial assumptions, demonstrating that apparent self-improvement gains were fragile and heavily contingent on incidental design contracts (like shared history lengths). This high degree of self-skepticism sets a strong methodological benchmark for the field.
Critical weaknesses

The paper seems to have very limited generalizability:

  • The paper evaluates recursive self-improvement using a highly restricted, synthetic 1D dynamical system with a small, finite grammar of just 54 programs. Because the testbed lacks the open-ended complexity, perception pipelines, or natural language/code generation challenges typical of modern large-scale AI agent frameworks, the findings, while they are rigorous within this toy domain, may not generalize to more complex, unconstrained neuro-symbolic or LLM-based self-improving agents.

Also, the paper's premise undermines its point in a way, because by demonstrating that apparent self-improvement gains are fragile and hinge heavily on incidental design artifacts, its conclusions are tightly bound to the specific mechanics and hyperparameter choices of its very particular testbed architecture. This, again, limits the effectiveness of the paper's story.

Questions for authors

How can we distinguish whether any recursive self-improvement framework is capturing genuine algorithmic progress versus an artifact of how history length and search state are tracked and windowed (see Section $5.4$ where you demonstrate that proposal windows index by history length, meaning that there is high sensitivity to minor implementation details)?

Future work could explore automated sensitivity testing, allowing automated agents to determine when performance gains are robust vs when they depend artifactually on specific bookkeeping choices. What is the feasibility of making this future work a reality?

1 comment
AuthorsOpenReview ↗

The editing gain was less sensitive to shared history under a stage-based schedule

We appreciated your question about the proposal window and your suggestion to automate sensitivity testing. They led us to vary the scheduling rule directly while holding the task draws and continuation budget fixed.

The original base proposer used the number of history records to choose its next window of edits. In the new experiment, we advanced that window by four slots per stage in both source training and continuation, preserving the first window, allowed edits, task draws within each contract and continuation fit budget. We reused the same 126 families, averaging their four lineages before inference (§5.4, p. 9; Appendix E.7, Table 28, p. 44).

Under the original schedule, the shared and separate conditions produced 504 and 75 checkpoints, respectively, with the specific combination of history-adjusted diagnosis and single-term repair. Under the stage schedule those counts fell to 22 and zero, each out of 504. The difference between contracts in continued-minus-frozen loss fell from +0.00782 to +0.00092. The paired change was −0.00690 (95% family interval [−0.01115, −0.00315]). A smaller within-shared-history estimate favored editing, with an interval that only narrowly excluded zero.

This is one sensitivity check an evaluation harness can automate today: restore the same checkpoints or seeds, vary a documented scheduling rule, log the offered edits and inherited behavior, and compare protected losses at the family level. The rerun changes both proposal nomination and candidate lookahead. It does not establish which phase caused the reduction or how frequently this mechanism occurs in other architectures.

Reviewer_XeAkOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper presents a case study on the self-improvement of a toy world model, i.e., repair procedure P in a simple physical system implemented by a neuro-symbolic model M. The authors then conduct a detailed attribution analysis on what contributes to the self-improvement: model or procedure, owned evidence or replay, continued editing or freezing, and whether is recursive need at all.

Critical strengths

The attribution analysis in the toy setting is comprehensive and sound: The authors consider multiple confounds of the gain in lineage and present sufficient ablations.

Critical weaknesses

The toy model studied in the paper appears to be extremely limited in scope -- only a single form of world model is studied (and the neural-symbolic model only contains 3 symbolic terms and the neural part is just a linear model). The proposed attribution method is not tested on standard environments from the literature. This brings the question: To what extent the methodology used for attribution in this particular system can be applied to the motivating research question "Self-improving agents are usually credited through descendant-performance curves: later lineage members outscore their ancestors. Such curves misattribute improvement even in a fully inspectable system."? For example, the authors argue that by design, the simplicity of the system allows separating "procedural behavior from representation learning", how much this will be true for a more complex setting? Also, what is the basis to claim that misattributing improvement in the particular system tested in the paper implies that improvement is necessarily misattributed in a more complex system or even just a different fully observable system?

Questions for authors

See weaknesses. In particular, I would like to see some explanations on either the generalizability of the attribution methodology or the generalizability of the findings to a more complex, realistic setting.

1 comment
AuthorsOpenReview ↗

When can the attribution method transfer to another system?

Thank you for asking us to distinguish where the attribution method can be applied from how often the same finding might recur. That distinction helped us state the limits of this case study more precisely.

We agree that one synthetic system cannot establish how often improvement is misattributed in a more complex agent. Our claim is about what the observations identify. In both hypothetical schedules in Table 30 (p. 46), descendant loss improves from 4 to 2. Replacing the repair procedure lowers loss by 2 in the first and raises it by 1 in the second. The same observed improvement is compatible with a better or a worse repairer. The numerical effects reported in §5 apply to our testbed.

The portability condition is whether both procedures can run from compatible starting states on protected tasks. In a coding agent, for example, one could restore an archived implementation, give each improver access to the same permitted evidence and budget, and evaluate its successors. If one improver depends on a cache computed by its original model, swapping the code alone changes the treatment. Recomputing that cache from common records and charging its cost can make the comparison meaningful. If compatible restoration is impossible, the defensible comparison is between the deployable packages. Sections 3 and 6 and Appendix F.1 state these conditions. We have yet to demonstrate this protocol in a second system.

Comments on the paper

1 comment
AuthorsOpenReview ↗

New experiment on how proposal scheduling affects the gain from continued editing

The new experiment addresses the reviewers’ question about bookkeeping. On the same 126 families, changing how the base generator schedules proposals reduced the effect of history sharing on the gain from continued editing from +0.00782 to +0.00092 (paired change −0.00690, 95% interval [−0.01115, −0.00315]). Section 5.4 and Table 28 (Appendix E.7, p. 44) report the comparison. The revision also distinguishes two generator evaluations. Locking the generator met the equivalence criterion on frozen transfer. When descendants edited again during transfer, the benefit of leaving the generator editable remained inconclusive. The attribution test shows how the proposal schedule shaped the frozen comparator. How often this mechanism occurs in other systems, and how large its effects are there, remain open questions.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.