acceptodds
Under review as a conference paper at ICLR 2027

EVOUNDO: RECOVERABILITY-CONSTRAINED SELF-EVOLUTION FOR LLM AGENT HARNESSES

Abstract

LLM agents increasingly modify their own prompts, tools, middleware, resources, and execution harnesses at runtime. Such self-evolution can improve capability, but a successful mutation may leave persistent effects that cannot be safely reversed in states different from the one in which it was created. We introduce EvoUndo, a framework for representing, synthesizing, diagnosing, and independently verifying recoverability of model-generated self-modifications across counterfactual states. Across 600 unseen one-shot self-evolution tasks, we identify 197 capability-improving mutations that fail recoverability verification. Under the original recovery representation, conventional repair strategies recover 0/197 of these natural failures. Deterministic oracle analysis recovers 48/197 under the original recovery language L0, while the extended recovery calculus increases empirical oracle recovery to 191/197. A protocollocked 2×2 grounding-by-expressivity intervention then separates two bottlenecks: exact state-address grounding increases successful recovery from 0/48 to 38/48 (79.2%) when the original language is sufficient, while extending the recovery language enables recovery on 142/143 (99.3%) failures in the oracle-defined S1 stratum. On the primary gpt-oss-120b backbone, adding exact-address diagnostics to the richer language reduces recovery to 133/143 (93.0%); a Qwen3.8-27B replication preserves the grounding and expressivity effects but not this negative interaction, indicating that the latter is model-dependent. These results indicate that reliable agent self-evolution requires co-designing verification, state grounding, witness semantics, and recovery-language expressivity rather than relying on iterative prompting alone.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
LLM agents, self-evolving agents, agent harnesses, recoverability, counterfactual verification, recovery synthesis, runtime safety, reversible agents

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_krPWOpenReview ↗
Rating
3(+2)
Confidence
2
Informativeness
2(+1)
Presentation
2(−1)
Soundness
2
Read review
Paper summary

EvoUndo tackles a practical problem for self-improving AI agents: how to undo changes to their own prompts, tools, and execution environment. The paper argues that useful modifications should come with a tested recovery procedure that works across different starting states, making recoverability part of the self-improvement process.

The framework captures relevant information before a modification, generates a recovery program, and checks whether running it restores the previous state. Failed checks produce feedback that guides further repairs. The original modification stays fixed throughout, preserving its useful behavior while the system improves the recovery logic.

Across 600 tasks, the authors identified 197 useful modifications that failed recovery checks. Conventional verifier-guided repair fixed none using the original recovery language. The experiments identified two main obstacles: missing information about affected states and missing operations needed to restore them. Better state-grounded feedback addressed the first, while a richer recovery language addressed the second. The best configuration repaired 180 failures, or 91.4%, raising overall admissible yield from 46.8% to 76.8% across the tested benchmark.

The main contribution is a framework that treats recovery as something agents must explicitly construct and verify. Reliable self-improvement depends on preserving state information, providing suitable recovery operations, and testing them together.

Critical strengths
  • The motivation is clear and the corresponding solution is systematic.
Critical weaknesses
  1. The evaluation does not establish when LLM-generated recovery offers an advantage over direct restoration. Effect-scoped snapshots outperform EvoUndo in both selective-undo settings: 100% versus 81% for different-surface changes, and 53% versus 43.7% for same-surface changes. There are no EvoUndo-only successes. Although the authors identify situations where direct restoration might be insufficient, those advantages are not demonstrated. This is my main concern: the experiments do not establish a compelling use case for the proposed synthesis machinery beyond the simpler baseline.

  2. The methodological novelty over prior undo frameworks is insufficiently established. GoEX [1] already synthesizes undo operations with an LLM and tests action–undo pairs in isolated environments, so generated and empirically checked recovery is not itself a new contribution. EvoUndo’s counterfactual verification and grounding–expressivity analysis provide a more specific contribution, but the paper needs to demonstrate what these additions achieve beyond the existing approach.

[1] Patil, Shishir G., et al. "Goex: Perspectives and designs towards a runtime for autonomous llm applications." arXiv preprint arXiv:2404.06921 (2024).

Questions for authors
  1. In what setting does synthesized recovery offer an advantage over direct restoration? Effect-scoped snapshots outperform EvoUndo in both selective-undo settings, with no EvoUndo-only successes.

  2. How much improvement remains when all repair methods have adequate recovery operations?

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

EvoUndo represents state capture and undo procedures in a typed language, tests whether modifications can be reversed across varied initial states, and uses verifier feedback to synthesize repairs. The experiments show that recovery depends on both diagnostic feedback and language expressivity. However, verification is empirical, operations are restricted to modeled harness state, and snapshots of affected state outperform EvoUndo in the evaluated settings. Its practical advantage for complex real-world actions remains unestablished.

Critical strengths

Supporting recoverability is important.

Critical weaknesses

More practical methods exist today.

Questions for authors

Checkpoint and restore exist in today's systems like containers, VMs, and file systems. Why rely on LLM synthesis when there are deterministic, efficient solutions?

Reviewer_r1ygOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
3(+1)
Soundness
3(+1)
Read review
Paper summary

This paper introduces EvoUndo, a framework for recoverability-constrained self-evolution of LLM agent harnesses. It couples forward mutations with pre-state witness capture, recovery programs, and typed effect contracts, and verifies recoverability through round-trip execution across counterfactual states. The authors investigate two recovery bottlenecks—diagnostic grounding and recovery-language expressivity—using a 2×2 factorial design. On 197 naturally occurring recovery failures, the best configuration recovers 180/197 (91.4%) cases. The paper further evaluates deterministic recovery oracles, snapshot-based baselines, cross-model replication, and fresh held-out tasks.

Critical strengths

S1. Important and well-motivated problem. The paper addresses recoverability as an important but relatively underexplored requirement for self-evolving agents, complementing existing work primarily focused on forward capability improvements.

S2. Systematic verification framework. Freezing forward mutations, independently tracking effect footprints, and verifying recovery across counterfactual states help prevent trivial solutions and incomplete recovery. The factorial design and paired statistical analyses are carefully organized.

S3. Transparent experimental reporting. The paper reports unfavorable findings, including the superiority of effect-scoped snapshots, non-replication of the negative diagnostic interaction, and limitations of the admission protocol. These analyses provide useful insights into the framework's boundaries.

Critical weaknesses

W1. The main experimental improvements are substantially confounded by recovery-language design. The base language L0 explicitly excludes recovery primitives for middleware, listeners, files, and sockets, while stratum S1 is defined by failures recoverable under L1 but not by the L0 oracle (Section 4.4). Consequently, the improvement from 0/143 to 142/143 on S1 (Table 2) largely reflects the introduction of previously unavailable operations rather than a novel recovery mechanism. Moreover, on S0, where L0 is already sufficient, D0L1 and D1L0 both achieve 38/48 recoveries, despite only the latter providing grounded diagnostics. This raises questions about whether the observed gains reflect diagnostic grounding, changes in the available operation vocabulary, or generation behavior. A failure-level breakdown of D0L0 and controlled comparisons with matched recovery primitives would help disentangle these factors.

W2. The necessity and practical advantage of LLM-based recovery synthesis remain unestablished. The deterministic constructive oracle recovers 191/197 failures without LLM inference by generating per-operation inverses in LIFO order (Appendix G). Although this oracle uses privileged information, its performance suggests that most evaluated recovery tasks admit straightforward deterministic compilation. More importantly, effect-scoped snapshots outperform EvoUndo on both selective-undo protocols (300/300 vs. 243/300 and 159/300 vs. 131/300), while requiring less storage and lower rollback latency (Appendix D.4, Table 5). There are no EvoUndo-only successes. The authors motivate their approach for settings where direct restoration is insufficient, but no experiment demonstrates such an advantage. A practical comparison against a deterministic inverse compiler and scenarios where snapshots or predefined inverses fail is needed to establish the unique value of LLM synthesis.

W3. Evaluation validity and generalizability remain limited. In the primary factorial study, aggregate hidden-set recovery statistics are returned during repair, so the hidden evaluation is not fully feedback-isolated (Section 3.2). On the fully isolated 300-task fresh holdout, recovery yield is only 35.0–51.0%, compared with 91.4% in the primary study (Section 7, Table 10). Although the protocols differ, this substantial gap deserves further explanation. Additionally, experiments use small simulated harness states without real agent frameworks, external side effects, or distributed dependencies, and the main configurations lack multi-seed replication. Finally, the negative diagnostic interaction (C3) involves only 11 discordant S1 tasks and does not replicate on Qwen3.8-27B under a different decoding configuration (Section 6.3, Appendix H.5). Its generality therefore remains uncertain.

Questions for authors

Given that the deterministic oracle recovers 191/197 failures without LLM calls and effect-scoped snapshots outperform EvoUndo in the evaluated settings, can the authors demonstrate a concrete class of harness mutations where deterministic inverse compilation and snapshots are insufficient, but LLM-based recovery synthesis provides a unique advantage?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.