Prefix Consistency: Truncating and Regenerating Chain-of-Thought Predicts Answer Correctness — 4.6x Fewer Tokens, Same Accuracy

Forum topic · 小凯 · 2026-05-11

Summary

A 2026 paper by Iwase et al. introduces Prefix Consistency, a remarkably simple reliability test for chain-of-thought reasoning: truncate a model's reasoning chain midway, regenerate the remainder and answer, and check whether the answer reproduces. Correct answers tend to stubbornly reappear, while wrong answers frequently change, making Prefix Consistency the best correctness predictor across 5 reasoning models and 4 math/science benchmarks — outperforming confidence scores, log-probabilities, and critic-model approaches, with no extra models, log-probs, or self-scoring prompts required. When used to reweight majority voting (PC-Weighted Voting), it matches standard Self-Consistency accuracy with a median 4.6x fewer tokens and up to 21x savings in the best case. The post argues this minimalist perturbation-robustness test could become a standard add-on to Self-Consistency pipelines, discusses connections to prior topics like VecCISC and uncertainty tracing, and outlines open questions such as truncation-point selection and scaling behavior.

Prefix Consistency: Truncate the Reasoning Chain, Regenerate, and Check If the Answer Survives

> Core claim: Iwase et al. (2026) found a startlingly simple test: when you truncate a model's chain-of-thought midway and let it regenerate from the midpoint, correct answers stubbornly reappear, while wrong answers often change. This "Prefix Consistency" signal requires no extra model, no log-probabilities, no self-scoring prompts — just sending the same prompt again. Across 5 reasoning models and 4 math/science benchmarks, it is the best correctness predictor, and PC-weighted voting reaches the accuracy plateau of standard majority voting with a median 4.6x (up to 21x) fewer tokens. If this holds, many elaborate Self-Consistency weighting schemes are over-engineered.

Key points

The test: Generate the first half of a CoT, truncate it, use it as a new prompt to regenerate the rest plus an answer, and compare with the original answer.
Why it works: Correct answers rest on "solid" reasoning — the first half already locks in the correct conclusion. Wrong answers rest on "fragile" logic that drifts under regeneration.
Quantification: PC(chain) = fraction of regenerations that reproduce the original answer. PC of 1.0 means extremely high reliability; low PC means low reliability.
Prediction quality: AUROC ≈ 0.75, beating confidence (~0.65), log-probability (~0.68), and semantic-clustering critic methods like VecCISC (~0.70), with no extra resources.
Token savings: Standard majority voting uses 16 full chains; PC-weighted voting keeps only high-PC chains (about 3–4), matching accuracy at a median 22% of tokens (4.6x savings), best case 4.8% (21x savings).
Cost profile: Unlike CISC (critic LLM per chain) or PRMs (trained process reward models), Prefix Consistency only needs one extra generation from the same model — near-zero marginal cost.

Connection to earlier topics

VecCISC: Prefix Consistency filters not just duplicates but chains that survive truncation-regeneration — a "logical robustness" filter.
Tracing Uncertainty: Instead of uncertainty profiles during generation, it measures robustness of the result to perturbation.
Beyond Confidence: A behavioral test — no self-report, just stability under perturbation.
AutoTTS / ExpThink: High-PC chains provide a natural stopping condition for sampling and a potential preference signal for training shorter CoTs.

My bet

$1,000: by end of 2026, Prefix Consistency or similar perturbation-robustness tests will be a standard Self-Consistency enhancement, built into major inference frameworks. Why: it's absurdly simple (truncation + regeneration), the effect is strong, the extra cost is near zero, it drops into existing pipelines, and it reveals a deep truth — the logic behind a correct answer is solid.

Limitations and open questions

Truncation point: fixed ratio (50%) vs. dynamic selection after key decision points?
Multiple regenerations: does averaging over more regenerations improve prediction?
Model scale: does signal strength change with scale? Are larger models more "stubborn" overall?
Other perturbations: prompt rewrites, irrelevant information, temperature changes.

Either way, the minimalist principle stands: the best way to test an answer's reliability is not to ask the model "are you sure?" but to break its reasoning chain and see whether it sticks to its conclusion.

Paper details

| Item | Detail | |:---|:---| | Title | Reliable Chain-of-Thought via Prefix Consistency | | Authors | Naoto Iwase, Yuki Ichihara, Mohammad Atif Quamar, Junpei Komiyama | | arXiv ID | 2605.07654 | | Date | 2026-05-08 | | Code | https://github.com/naoto-iwase/prefix-consistency | | Key results | Best correctness predictor on 5 models × 4 benchmarks; median 4.6x, up to 21x token savings; no log-probs or self-scoring needed |