Large Language Models Cannot Self-Correct Reasoning Yet

Self-correction - asking a model to check and fix its own output - is widely assumed to improve reasoning without any external feedback or retraining.

Not a mechanism but a CONTROL EXPERIMENT. Critically examines INTRINSIC self-correction, where the LLM attempts to correct its initial response based solely on its inherent capabilities, without external feedback, across model families, benchmarks, and prompt formulations.

None, and that is the finding: LLMs struggle to self-correct without external feedback, and at TIMES PERFORMANCE EVEN DEGRADES after self-correction. Not recorded as a null but as the paper's negative result, which is load-bearing - it means self-correction loops are a pure COST with no expected accuracy return in the intrinsic setting. Any serving system implementing a self-correction retry loop is paying output tokens for tokens that measurably do not help.

ADOPTION of the FINDING, not a technique: this paper is in production in the sense that its negative result shapes deployed designs - systems that ship self-correction retries do so with external verifiers or accept the cost knowingly. VENUE: ICLR 2024 per the arXiv comment field, so venue rests on the primary source. INCLUDED DELIBERATELY as the counterweight to the rest of the inference-time-compute sub-area: chain-of-thought, self-consistency, and budget forcing all spend output tokens expecting accuracy back, and this paper is the evidence that at least one common token-spending strategy does not deliver. The distinction it draws - intrinsic self-correction versus correction WITH EXTERNAL FEEDBACK - is the operative one, and a reader who quotes this as 'self-correction does not work' has over-read it; process supervision with a verifier (pap-eff-verify-step-by-step) is the external-feedback case that does work. AFFILIATIONS FILLED 2026-10-04 (paper-verification pass): from the Semantic Scholar graph API (paper/arXiv:2310.01798), which is the only machine-readable source that carries author affiliations — the arXiv Atom feed returns author names only, and no ORCID is exposed by either API. One known upstream typo corrected: "Tel Aviv Univesity" -> "Tel Aviv University". S2 affiliation coverage is partial (120 of 249 corpus records carry any), so these lists may be incomplete rather than exhaustive. VENUE_TRACK BACKFILL 2026-10-04 (paper-fields pass, agent w3p): venue_track 'main'. This is STRUCTURAL rather than researched: iclr has no Findings volume and no workshop track in which a paper could appear, so 'main' is the only possible honest value and no per-paper evidence was needed. The field is annotated here so that 'findings' and 'workshop' remain MEANINGFUL - a filter for findings=true must return only the papers that actually are findings, which requires every other paper to be explicitly NOT findings.