Research map / Large language models / Reasoning and chain of thought

Reliable Chain-of-Thought via Prefix Consistency

NeurIPS 2026 · Large language models › Reasoning and chain of thought · approach: Long CoT dynamics and efficiency · 0 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Proposes prefix consistency, a reliability signal from truncating CoT traces and regenerating continuations, used to weight majority voting (PC-WMV) without log-probabilities.

Problem

Standard majority voting (self-consistency) fails when the correct answer is a minority among sampled reasoning traces, and existing weighted voting signals (log-probs, self-certainty, verbalized or P(True) confidence) poorly separate correct from wrong traces on hard problems.

Approach

Truncate each CoT at a fraction tau, regenerate the continuation, and score each candidate answer by how often it reappears in the group; correct traces reproduce their answer more often than wrong ones, so this reproducibility serves as a log-prob-free weight for voting.

Key results

• Prefix consistency has the best macro-averaged AUROC in 15 of 20 (model, benchmark) cells, typically around 0.7.
• PC-cubic reaches Standard MV plateau at a median 4.6x fewer tokens, and up to 21x fewer than Standard MV across settings.
• The discrimination gap D = r_C - r_W is positive on all 20 model-benchmark cells, e.g. 43.7 points for GPT-OSS-120B on FrontierScience-Olympiad.

Surprising findings

• Baseline confidence signals such as DeepConf tail and P(True) are near random (AUROC about 0.5) on harder cells.
• PC-WMV does not beat Standard MV on cells where Pass@1 is already close to the Standard MV plateau.

Limitations

• Requires an explicit visible CoT that can be truncated, so it does not apply to closed models with hidden reasoning.
• Additional inference cost from regeneration tokens; hyperparameters tau, K and w fixed rather than tuned.
• Evaluation limited to reasoning models and math/science benchmarks; mechanism behind r_C > r_W not explained.

Experimental setup

Datasets: FrontierScience-Olympiad HMMT Feb 2026 AIME 2025 Brumo 2025

Benchmarks: FrontierScience-Olympiad HMMT Feb 2026 AIME 2025 Brumo 2025

Metrics: macro-averaged AUROC accuracy at fixed token budget token-efficiency ratio discrimination gap D

Baselines: Standard MV Self-certainty DeepConf tail P(True) Adaptive Consistency Early-Stopping Self-Consistency Verbal 0-100 DeepConf bottom-10%

Builds on: self-consistency DeepConf Adaptive Consistency Early-Stopping Self-Consistency

Closest papers by meaning

• Optimal Self-Consistency for Efficient Reasoning with Large Language Models (ICML 2026 · similarity 0.91)
• Pause and Reflect: Conformal Aggregation for Chain-of-Thought Reasoning (NeurIPS 2026 · similarity 0.89)
• THE PATH OF LEAST RESISTANCE: GUIDING LLM REASONING TRAJECTORIES WITH PREFIX CONSENSUS (ICLR 2026 · similarity 0.89)
• CITE: Anytime Valid Statistical Inference in LLM Self-Consistency (NeurIPS 2026 · similarity 0.89)
• Latent Self-Consistency for Reliable Majority-Set Selection in Short- and Long-Answer Reasoning (AAAI 2026 · similarity 0.88)
• FaithCoT-Bench: Benchmarking Instance-Level Faithfulness of Chain-of-Thought Reasoning (ICLR 2026 · similarity 0.88)