Can language models fix their own reasoning mistakes?

The paper defines "intrinsic self-correction" as an LLM revising its own answer "based solely on its inherent capabilities, without the crutch of external feedback," and tests it on reasoning with GPT-3.5-Turbo, GPT-4, GPT-4-Turbo, and Llama-2-70b-chat. The finding: "LLMs struggle to self-correct their responses without external feedback, and at times, their performance even degrades after self-correction." Prior papers claiming gains (Kim et al. 2023; Shinn et al. 2023) turn out to depend on oracle labels telling the model whether its answer was already right — "the improvements vanish when oracle labels are not available." A second artifact is prompt design: some claimed self-correction gains "stem from the sub-optimal prompt for generating initial responses," where the feedback step merely supplies instructions the first prompt should have included; folding those instructions into the initial prompt erases the advantage.

The paper's mechanism, in its own terms: when the model is already well-aligned and the initial prompt well-designed, "the initial response should already be optimal relative to the prompt and the specific decoding algorithm." A feedback prompt is then an additional, unhelpful input that can "bias the model away from producing an optimal response to the initial prompt." The paper also re-examines multi-agent debate (Du et al. 2023), where multiple LLM instances critique each other's answers, and replicates it on GSM8K at matched inference cost against self-consistency. Debate's gains are "no better than self-consistency... when considering an equivalent number of responses" — the paper argues debate is really "a means to achieve 'consistency' across multiple model generations," differing from self-consistency only in whether the vote is model-driven or count-based, so "the observed improvement is evidently not attributed to 'self-correction,' but rather to 'self-consistency.'"

This sharpens Is reflection in reasoning models actually fixing mistakes? and Does reflection in reasoning models actually correct errors?: those studies found reflection in trained reasoning models rarely overturns the initial answer; this paper gets the same directional result via a different mechanism — explicit multi-round self-correction prompting on non-reasoning-tuned chat models — and adds the methodological diagnosis (oracle-label leakage, unfair baselines) explaining why earlier work saw gains where none existed. It also parallels Does self-revision actually improve reasoning in language models? in showing degradation rather than improvement from revision, and gives a prompting-level counterpart to What limits how much models can improve themselves?: without an external verifier (oracle label, tool, or trained critic), the model has no advantage over its own generation to exploit, so correction has nothing to work with. Why does self-correction training on offline data fail? addresses the same gap at training time rather than at inference time.

The excerpt tests reasoning tasks only, on 2023-era chat models prompted explicitly to self-correct — it does not establish that self-correction fails for other domains; the authors note style, safety, and preference alignment are reported elsewhere as cases where LLMs "properly evaluate whether a response is inappropriate," unlike judging their own reasoning errors. It also does not test whether models trained specifically to produce useful reflection (rather than prompted post hoc) behave differently, leaving open whether the limitation is about prompting LLMs to self-correct or about self-correction as such.