EvoUndo: Making Agent Self-Modifications Recoverable, Not Just Capable

EvoUndo introduces a framework that verifies recoverability of model-generated self-modifications across counterfactual states. In 600 one-shot tasks, 197 capability-improving mutations failed recoverability; conventional repair recovered none, while an extended recovery calculus recovered 191/197 in oracle analysis. Grounding and expressivity interventions isolate two bottlenecks, but a negative interaction on gpt-oss-120b is model-dependent.

September 1, 20261 source65% confidence

75Strategic

Strategic score

Evidence strength72

Technical novelty78

Commercial potential55

EvoUndo makes agent self-evolution safe by changing what counts as a successful modification: a mutation is not accepted merely because it improves capability, but because it can be independently verified as recoverable across counterfactual states. This is a shift from capability-constrained evolution to recoverability-constrained evolution. The new work reports that in 600 unseen one-shot self-evolution tasks, 197 capability-improving mutations failed recoverability verification, and that conventional repair strategies recovered none of them under the original recovery representation.¹

Context

The problem is that a successful self-modification may persist in a mutated harness and be impossible to reverse in states different from the one in which it was created. EvoUndo formalizes this as a recoverability property and separates two bottlenecks: exact state-address grounding and recovery-language expressivity. In the intervention, exact-address grounding recovered 38/48 (79.2%) of oracle recoverable failures when the original language was sufficient, and extending the recovery language recovered 142/143 (99.3%) of oracle-defined S1 failures.¹

Assessment

The evidence for the bottleneck separation is strong because it is built on a deterministic oracle and a protocol-locked 2x2 design. It is, however, only one study: the main negative interaction—adding exact-address diagnostics to the richer language reduced recovery on gpt-oss-120b from 142/143 to 133/143 (93.0%)—was not reproduced on Qwen3.8-27B, so it is model-dependent. The abstract does not report the task distribution, oracle definition, or repair baselines, which limits independent assessment.¹

Caveats

What would change our mind: if the independent verification turns out to be coupled to the model that generated the mutation; if the oracle-defined strata leak evaluation signal; or if the recoverability failures are rare or easy in deployed settings. Under those conditions the constraint shift would not generalize. For now, the contribution is a concrete, testable proposal that co-designs verification, state grounding, witness semantics, and recovery-language expressivity.¹

Source text

Hugging Face Daily Papers — EvoUndo: Recoverability-Constrained Self-Evolution for LLM Agent Harnesses · August 28, 2026

Computer Science > Artificial Intelligence

Title: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 protocol-locked 2x2 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.

Access Paper:

Bookmark

Bibliographic and Citation Tools

Code, Data and Media Associated with this Article

Demos

Recommenders and Search Tools

arXivLabs: experimental projects with community collaborators

arXivLabs is a framework that allows collaborators to develop and share new arXiv features directly on our website.

Both individuals and organizations that work with arXivLabs have embraced and accepted our values of openness, community, excellence, and user data privacy. arXiv is committed to these values and only works with partners that adhere to them.

Have an idea for a project that will add value for arXiv's community? Learn more about arXivLabs.

The source text above remains the property of its publisher and is reproduced for commentary.

↑ Back to top