Self-Consistency Improves Chain of Thought Reasoning in Language Models

Chain-of-thought prompting still uses NAIVE GREEDY decoding, which commits to the first sampled reasoning path even when several paths would have reached the right answer.

DECODING STRATEGY, not a model change. Sample a diverse set of reasoning paths instead of only the greedy one, then select the most CONSISTENT answer by marginalising out the sampled reasoning paths - i.e. aggregate the final answers across sampled paths and take the majority. Exploits the intuition that a complex reasoning problem admits multiple ways of thinking leading to one correct answer.

Not a latency paper - it is a deliberate ACCURACY-FOR-COMPUTE TRADE, and the trade is the point. Accuracy gains: GSM8K +17.9%, SVAMP +11.0%, AQuA +12.2%, StrategyQA +6.4%, ARC-challenge +3.9% over greedy chain-of-thought. The compute side is unquantified in the abstract because it depends entirely on the number of sampled paths, which the operator chooses. COST IMPLICATION recorded explicitly: self-consistency multiplies output tokens by the sample count, and since decoding is bandwidth-bound the latency is close to LINEAR in sample count - so this is the cheapest-to-implement and most expensive-to-serve technique in this sub-area, with no engine-level mitigation.

ADOPTION: in-production in the sense that self-consistency is a decoding parameter any server exposes (num_samples / n>1) and needs no code changes; whether a given deployment enables it is an application choice. VENUE: ICLR 2023 - corroborated by the arXiv comment field, which explicitly says 'Published at ICLR 2023 ... v4: camera ready version at ICLR 2023', so this venue claim rests on the primary source rather than secondary aggregation. THE STRUCTURAL POINT FOR THIS REPO: together with chain-of-thought (pap-eff-chain-of-thought-prompting) and budget forcing (pap-eff-s1-budget-forcing), this record is the clearest statement that inference-time compute is bought in OUTPUT TOKENS, which are the bandwidth-bound quantity. Any system-level speedup for reasoning models should therefore be reported against a stated token count, not against a stated wall-clock time alone. DANGLING REF STILL OPEN (2026-10-04 paper-read pass; supersedes the earlier 'LEFT UNRESOLVED' note, which reached the same conclusion but only from this record's own text - this time the PAPER itself was read). No A100 and no NVIDIA GPU anywhere in the paper; every experiment it describes ran on TPUs or an API: "For UL2 we use TPU v3 (2x2 configuration, 4 chips, 8 cores). For GPT-3 models the experiments are done though the public API. For LaMDA-137B we use TPU v3 (8x8 configuration, 64 chips, 128 cores). For PaLM-540B we use TPU v4 (4x4x12 configuration, 192 chips, 384 cores)" (Appendix A.2 'Details on resources and inference'). Source: https://arxiv.org/pdf/2203.11171 . Left dangling. This paper predates the A100 in its evaluation entirely, so the ref is not merely ambiguous, it is unsupported. AFFILIATIONS FILLED 2026-10-04 (paper-verification pass): from the Semantic Scholar graph API (paper/arXiv:2203.11171), 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.