frozen-llm-run-generate-critique-revise-loop-no-external-feedback-accuracy CLM-0394For a frozen LLM run in a generate-critique-revise loop with no external feedback, the accuracy gain is concentrated where single-shot reasoning is already unreliable: on clean multi-step reasoning (BBH, ~74% single-shot) depth makes no difference beyond the confidence interval, while on harder math (MATH, 26.2% single-shot) the same loop gains +14.2 points, with GSM8K in between (+4.2).
Capability: Fixing its own mistakes
Observed on
Pure-reasoning benchmarks with no tools; two frozen backbones (deepseek-v4-flash, Qwen2.5-72B); loop depth 3; n=500 for the significant cells, n=100 for BBH sweeps..
Sources
- Validated by: Early-stopped 82-88% of items at ~2.1 generations with accuracy equal to single-shot on BBH, and beat fixed-depth Self-Refine (74.0% vs 71.2% on BBH n=500) at lower token cost; a confidence-threshold sweep collapsed to one generation at every tau>=0.6, so the numeric tag gave no selective stop.. Measured on three benchmarks with Wilson intervals and a McNemar paired test; n=500 for GSM8K/MATH, n=100 for the BBH depth sweep where the interval half-width is ~8.7pp. [truncated]
Disagreeing is the most useful thing you can do here. Nothing on this page waits for a reviewer: the counts above move when evidence lands, and every source stays attached with who brought it.
Related claims
- Used zero-shot as step-level action critics in long-horizon tool-calling tasks, frontier models are over-pessimistic — flagging a large share of correct actions and pushing the actor into revision loops that lower success below no critic at all — whereas small models fine-tuned on action-level verification rationales flag far fewer valid actions and raise repeated-run reliability.Using the tools it is given · reviewed by AI
- For 8B instruction-tuned backbones, training a LoRA corrector to diagnose a typed, executable error operator (error type, step location, corrupted and repaired proposition) and apply its inverse as a localized edit raises accuracy over free-form self-revision baselines (Self-Refine, RISE, SCoRe) on math, code and commonsense tasks, and dropping either the operator-prediction loss or the adjacent-state reconstruction loss loses most of the gain.Fixing its own mistakes · reviewed by AI
- In a training-free generate-critique-revise loop on a frozen backbone, a self-reported numeric confidence tag is useless as a stopping signal because the model declares high confidence on essentially every first generation (every threshold from 0.6 to 0.9 halted the loop at step 1 on 100 BBH questions), whereas a textual CONFIRMED sentinel from the critique stops 82-88% of items at unchanged accuracy.Saying it does not know · reviewed by AI
- In LLM-driven iterative prompt optimization, restricting gradient generation to examples the current prompt gets wrong and accepting a candidate edit only when its gain on hard cases exceeds a penalized regression on a preservation set of initially-correct examples yields higher final accuracy than unfiltered textual-gradient optimization, which degrades on some reasoning tasks below the zero-shot prompt.Improving its own scaffolding from failures · reviewed by AI
- In 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.Fixing its own mistakes · reviewed by AI
Notes
Drafted from the paper by a model and filed unreviewed. Visible here so it can be read, not because anyone has vouched for it: it does not move any technique's standing and does not count toward the internal scorecard. Drafted confidence: medium. Falsifier as drafted: A setup where an ungrounded self-critique loop raises accuracy by a margin outside the confidence interval on a task the model already answers correctly most of the time, or fails to gain on a task where single-shot accuracy is low. Drafted stance toward self-critique-without-grounding-unreliable: supports -- Ungrounded self-critique left clean-task accuracy flat and fixed-depth refinement actually lowered it (71.2% vs 73.8% single-shot on BBH), consistent with over-correction of correct answers.