self-repair · activeFixing its own mistakes
Whether and how a model can diagnose and correct a wrong answer once it's already made.
Tags: agentic, coding-agent, math
This topic covers self-correction and self-critique behavior: does reviewing or revising an answer make it better or worse, and what turns out to matter is what's driving the revision — the model's own unaided judgment, or a concrete external signal like a failing test. The two are easy to conflate under one banner ("self-correction") and behave very differently, which is exactly the kind of thing a bare "capability score" would have hidden.
Claims
- observationsingle paperWith weights frozen and no stronger model in the loop, an agent can raise its own pass rate by rewriting the scaffolding it runs inside — clustering its failures, proposing minimal edits to prompts, tools and policy, and keeping only edits that improve one split without degrading another: all 9 of 9 model-by-benchmark pairs improved, by up to 40.6 percentage points (GLM-5 on AppWorld, 44.4% to 85.0%), with weak models gaining most; but the promotion gate reads the held-out split, so the reported held-out gain is not clean generalisation, and there is no matched-compute baseline.
- mechanismmechanism reasoningIn an agent's verify-and-fix loop, deterministic checks — tests, linters, schema validators — should be exhausted before any model-based judge is added, because they are free, fast and give the same verdict every time, while an LLM-as-judge costs tokens per run, returns non-deterministic verdicts, and so belongs only on properties no rule can express, and even then as an advisory signal until measured against human review.
- mechanismreplicatedcontestedReflect-and-retry raises task success when the feedback in the loop is a genuine external signal — a failing test, a compiler error, an environment outcome — but not when the "feedback" is the model's own unaided critique. The grounding, not the reflection step, is what does the work.
- mechanismsingle paperWhen a model gets a real environment signal after acting — a test result, a tool error, a task-success indicator — reflecting on that signal in words and retrying substantially improves success rates on multi-step coding and decision-making tasks over a single attempt.
- mechanismsingle paperIn multi-step visual QA, a critic step run by the same model improves final accuracy only when it is shown the candidate evidence alone and denied the generator's textual rationale; conditioned on that rationale, the critic ratifies the generator's initial hypothesis and the loop performs worse than no critic at all.
- observationsingle paperIn open-ended editing tasks with no unique correct output (multi-slide design editing), a separate judge model that scores the structural diff between the original and edited artifact against the instruction, and returns its critique verbatim as the next-turn instruction, raises instruction-following over a single pass, but the gain is selective: most tasks pass on the first attempt, and a minority of the ones that enter the loop get worse unless an earlier iteration is restored when the judge's own score declines.
- mechanismsingle paperWithout an external, ground-truth signal — a failing test, a compiler error, a verifier's output — a model's own critique of its reasoning is not a reliable improvement signal, and asking it to review and revise a correct answer often turns it into a wrong one.
- mechanismsingle papercontestedPrompting a model to generate its own feedback on a draft output and revise accordingly improves quality on open-ended generation tasks (dialogue, code, review-writing) over a single-shot attempt.
- mechanismmechanism reasoningThe agent that produced an artifact is a biased judge of it — it holds the context and the incentives that skew its assessment — so verification belongs with a deterministic sensor or a separate verifier that reports failures back rather than rewriting the output, and in a multi-agent system the verifier is the one component no agent may override.
- mechanismsingle paperTool-augmented multimodal models trained only on how to call vision tools accept tool outputs uncritically; adding supervised trajectories where tool evidence is checked against the image and re-queried, plus a reward for correction that reaches a correct answer, raises accuracy on perception-heavy benchmarks (counting, depth, spatial relations) in 3B and 7B Qwen2.5-VL, while slightly lowering it on knowledge-heavy ones.
Techniques
- Deterministic checks before LLM judgesprocessWire linters, test suites and schema validators into the agent's loop before adding any model-based judgment.
- Let the agent revise its own harness, regression-gatedprocessCluster failed traces into signatures, have the same fixed model propose a few minimal harness edits, and promote only edits that improve one split without degrading another.
- Repair with external feedbackprocessGive the model a concrete external signal, such as test output, and let it reflect and retry.
- Separate the verifier from the producerprocessThe agent that produced an artifact does not judge it; a deterministic sensor or a separate verifier agent does, and reports rather than rewrites.
Capabilities are a way of carving up the subject, and carvings are arguable. Say so if this one is wrong — especially a proposed one, which a pipeline added because several papers used the same framing, not because anyone decided it was right.