acceptodds
Under review as a conference paper at ICLR 2027

Can Self-Evolving Agents Recover from Their Own Mistakes?

Abstract

Self-evolving agents modify their own prompts, workflows and parameters and select modifications by a score. Score-based selection guards against modifications that lower the score, but not against those that remove something the agent needs while the score stays high. We ask whether such loops can recover from these silent mistakes. Without its archive-based parent selection, a stock workflow optimizer deletes its agent's input placeholders, turns every exception into a graceful exit, and replaces its model call with a nonexistent method; the agent never calls a language model, yet its validation score equals the initial agent's, because a third of validation tasks pass when the agent does nothing. Two other optimizers fail differently; only the archive protects the deployed version. In a controlled IFEval-style task with fifteen verifiable conventions, reporting how many conventions an output satisfies prevents the losses that pass/fail feedback causes; but once a convention is lost, neither counts nor anonymous identifiers of the failing convention restore it in any of four rewriters, while supplying its statement does. We prove that feedback which only evaluates the current state cannot repair a loss faster than the rewriter can guess the lost content. An unseen tool outage is diagnosed only when error traces reach the optimizer, and the self-written exception handler turns the fault's score drop into a score gain. In self-training, a small model collapses to a constant answer that an accuracy gate cannot see. We argue for evaluating self-evolving systems by whether their mistakes remain detectable and recoverable.

Primary area:
generative models
Keywords:
Self-evolving Failure Mode

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_HuXmOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

The paper aims to investigate the failure modes of self-evolving agents while they repeatedly revise their own harness/workflow or their own weights during online interaction with environments with different styles of feedback. The paper's goal is also to test whether self-revising agents can retain information needed to find, reverse, or repair their own mistakes. The paper tests a few settings. The first one is a workflow revision based on the EvoAgentX implementation of AFlow on the tau^2 Bench. They tested the model revising its own harness while receiving feedback from the environment. They show that the agent can make significant failures when revising their own workflow, causing catastrophic degradation in the agent's real functionality, but the validation score remains unchanged since doing nothing passes some of the validation tasks. This happens only under the accept-all modification workflow, not the stock archive-based parent selection workflow. The second setting is on prompt engineering inspired by IFEval and compares different types of feedback. The paper tries to show that knowing what is wrong does not mean the model knows how to fix the requirement. The third task is HTTP 503 failures assigned with some probability to a tool used by the agent. The authors show that with traces, the models can identify the problems, but without traces the models fail to correct that. The last task is using STaR-style self-training with LoRA. The paper shows that with repeated training, the Qwen 3 0.6B model would converge to selecting almost all "A" on MedMCQA.

Critical strengths

Self-evolving agents may fail even when validation scores indicate similar levels. This is the main point the paper tries to show and experiments are built around this point.

Critical weaknesses

Although the paper presents examples of degradation that monitored scores fail to expose, the experiments do not sufficiently establish how broadly these failures occur or whether they reflect a common mechanism of impaired recovery following self-modification.

  1. The workflow experiments do not adequately isolate loss of recoverability caused by self-modification from limitations of the feedback and harness design. My concern here is, for instance, if the agent can assess a history of its past revisions and failure reports, then it should be granted this skill otherwise, it feels more of a harness problem to me. In the tool-outage experiment, providing error traces enables the optimizers to identify the fault, whereas score-only feedback does not. Successful diagnosis does not guarantee repair, but this comparison primarily establishes the importance of making diagnostic information available. As acknowledged in Appendix D, it does not demonstrate a previously functioning diagnostic channel being destroyed by self-modification: fault detection is the same for the original and modified checkpoints under each feedback condition. Furthermore, the headline workflow collapse occurs after replacing AFlow’s archive-based selection with an accept-all policy, while the stock selected workflow remains functional. This means that validating before applying fixes is important but does not mean the model cannot fix the issues with good context management or feedback. These results leave unclear how much the observed failure generalizes to systems retaining standard selection and diagnostic mechanisms.
  2. The parameter-update experiments provide limited evidence about recovery from self-induced damage. The STaR-style experiments demonstrate that correctness-filtered training can cause Qwen3-0.6B to converge toward a constant answer while accuracy remains near the majority-answer baseline. However, they primarily examine the emergence and prevention of collapse, rather than whether an already-collapsed model can recover under different corrective interventions. What if we consider a simple workflow: we do checkpointing and prevent the model from collapsing by reverting the training if we find the validation scores on answer choices uniformly distributed validation questions drop or the answer converges to a single choice? Their connection to the paper’s central recovery question is therefore less direct than that of the prompt experiments. Moreover, Llama-3.2-3B does not collapse under the tested training intensity, suggesting that the phenomenon depends on the model’s initial competence and its distance from the majority-answer baseline. The manuscript should clarify these boundary conditions and explain which conclusions about recoverability the parameter experiments independently support.
Questions for authors
  1. What are the evidences you find most convincing that shows self-modification or workflow revision reduces subsequent recoverability with a better harness design, if any?
  2. What conclusion about recovery is supported by the STaR self-training experiment?

These are questions that I have for the authors. However, I find it might be hard for me to further improve the scores unless new experiments or insights are added.

Reviewer_oF9ZOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
1
Soundness
2
Read review
Paper summary

The paper builds on an observation that self-improving methods can lose some functionalities without decreasing the benchmark performance, hence leading to a collapse in a setup where the acceptance of the modification is based on the score. They provide examples of such collapses, basic analysis, and prove that in such a setting, after losing an instruction, the model is not able to recover it faster than guessing it as if it were never there.

Critical strengths
  • The paper is novel in the sense that it is analysis/theory oriented paper in the agentic RSI field.
  • The problem of collapse in RSI is a significant issue
  • The silent collapse is an interesting observation that, with its significance sufficiently justified, could be a good point of analysis for future RSI systems.
Critical weaknesses
  • For an analysis paper, the provided quantitative analysis and experiments is substantially insufficient. The results are mainly limited to a use case study of a selected model. There is no evaluation of multiple models. The model does not provide any insight into how often such phenomenon happens.
  • The collapse is a known problem and is addressed by having an archive of agents proposed in DGM[1]. The paper mentions DGM and mentions that the problem might still appear there, as the selection is based on the score, yet it doesn't provide any analysis.
  • The paper ignores methods that are selecting the agents based on different metrics than score (such as HGM[2]) and does not even mention them ( first sentence of the abstract: "Self-evolving agents modify their own prompts, workflows and parameters and select modifications by a score.), yet the observations of the paper should suggest that these should handle silent collapse (as defined by the authors) much better than the score-based methods. Yet the paper does not provide any analysis of the archive-based systems.
  • The writing in the paper is poor. Lot of jumps between contexts inside a paragraph without proper connecting sentences.

[1] Darwin Godel Machine: Open-Ended Evolution of Self-Improving Agents (ICLR 2026) [2] Huxley-Gödel Machine: Human-Level Coding Agent Development by an Approximation of the Optimal Self-Improving Machine (ICLR 2026)

Questions for authors

In general i like the motivation to analyse the failure mode of self-improving systems due to silent collapse. However, in order for the paper to be useful to researchers, it needs to analyse this in significantly more depth. I would like to see:

  • a quantitative analysis of silent collapse across multiple methods. How often it happens and how often it leads to "loud collapse" or it mostly stays flat afterwards?
  • Analysis of the archive-based self-improvement methods that are benchmark score-based (such as DGM) and their silent collapse, what is the frequency of silent collapse there
  • Analysis of the archive-based self-improvement methods that are not benchmark score-based (such as HGM) and their silent collapse. If the authors' claim is right, the node that collapsed silently should be selected significantly less often in HGM than DGM (because the averaged-tree score selection naturally gives a bias to an uncollapsed parent). If the authors could show that, this would be a good alternative explanation of the performance boost of HGM over DGM.

At the current state the paper is not ready for the conference. However, I like the motivation and the idea and I am open to change the score if provided with sufficient additional analysis.

Reviewer_rt7qOpenReview ↗
Rating
2
Confidence
1
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

The paper investigates the self0evolving agents can recover from their own mistakes and presents several ideas and findings. The paper is hard to read for one not in the exact area. It may be much more approachable if some discussion at the intuition level can be added.

Critical strengths

Whether and how self-evolving agents may recover from their own mistakes is an important and interesting problem. The paper presents several observations and ideas. Some insights, such as preventing a loss is not repairing it, are critical.

Critical weaknesses

The paper is hard to read. For example, the abstract is not written for a general reader in the ICLR community. It is filled with a lot of details without an introduction to the problem and the intuition and insights of the solutions. As another example, the paper can use a simple example to illustrate "The harder case is a modification that removes something the agent needs while the score stays high, or even rises."

As I am not in the area, I don't have much insights into the technical parts.

Questions for authors

Make sure that the paper is approachable for general ICLR audience, including those who do not work on self-evolving agents.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.