Optimal Self-Consistency for Efficient Reasoning with Large Language Models

AI summary

Authors: Austin Feng (Yale University; work done during an internship at Noah's Ark Lab), Marius Alonso (Noah's Ark Lab, Paris), Ambroise Odonnat (Noah's Ark Lab, Paris; Inria, Rennes), Vasilii Feofanov (Noah's Ark Lab, Paris; 42.com, Amsterdam), Ievgen Redko (Noah's Ark Lab, Paris) arXiv: 2511.12309v2 [cs.LG], 30 Jun 2026 · License: CC BY-NC-ND 4.0

Overview

Research area: Test-time inference for large language model reasoning, combining mode estimation, voting theory, and statistical learning theory. Technical level: Advanced. The paper builds on exponential error bounds, martingale confidence sequences, Laplace transforms, and constrained convex optimization. Scope: A theoretical and empirical study of the sample efficiency and scaling laws of self-consistency (SC) and its fixed- and dynamic-allocation variants, culminating in a hyperparameter-free adaptive algorithm called Blend-ASC.

What This Paper Is About

Self-consistency is the common practice of sampling many chain-of-thought answers from an LLM and returning the most frequent one, which is equivalent to a plurality vote and to estimating the empirical mode of the model's answer distribution. It works, but applying it naively across a whole dataset is prohibitively expensive, and there has been no unified theory explaining how quickly it converges or how samples should be spent. This paper supplies that theory, derives scaling laws, and uses the results to build a sampling scheme that reaches the same performance with far fewer samples.

Key Contributions

1. Theoretical analysis. The paper provides a first complete study establishing test-time scaling laws for vanilla SC, fixed-allocation SC, and dynamic-allocation SC. The per-question bound is stated to be tighter than prior work on per-question scaling (Huang et al., 2026), and it is the first analysis to cover both fixed and dynamic allocation.
2. Empirical validation. The theoretical results are checked using three recent LLMs of varying sizes on three popular reasoning benchmarks, demonstrating real-world applicability. The paper also validates the predicted power-law (and faster) error decay on benchmark data.
3. Theory-informed SC (Blend-ASC). Motivated by the analysis, the authors introduce Blend-ASC, a self-consistency variant that dynamically allocates samples to questions during inference and is described as achieving state-of-the-art sample efficiency for a given budget, without relying on expensive hyperparameter tuning.
4. Practical usability. In contrast to existing variants, Blend-ASC is reported to be hyperparameter-free, to support batching, and to fit any budget of samples, making it easy to plug into existing self-consistency pipelines.

Main Findings

• Self-consistency is mode estimation. SC returns the empirical mode of the answer distribution; a question is "aligned" if the model's distribution has the true answer as its mode. Sample efficiency is therefore the rate of convergence to that mode, and the problem can be recast as i.i.d. voters choosing among candidates.
• Margin governs the decay rate. For an aligned question, the reducible error satisfies err(x,q) ≤ exp(-x((√p₁ − √p₂)² + ε)), where p₁ ≥ p₂ ≥ p₃ ≥ … are the sorted answer probabilities, ε → 0 as x → ∞ at rate O(log x / x), and the low-probability tail is grouped so that only the top k ≪ K answers matter. The rate m = (√p₁ − √p₂)² is called the margin.
• The bound is tighter than prior work. Huang et al. (2026) require x ≥ 2 log(1/δ)/(p₁ − p₂)² samples for error δ; this paper needs x ≥ log(1/δ)/((√p₁ − √p₂)² + ε). The paper argues margin (√p₁ − √p₂)² is a more natural confidence measure than the absolute gap p₁ − p₂.
• Margin correlates with empirical decay. Across Llama-3.2-3B, Qwen-2.5-Math and Qwen-2.5-32B on MMLU, MATH and GPQA-Diamond, margins estimated from 100 samples track the fitted exponential decay rate of the mode-estimation error (decay fit for x ≥ 16, and x ≥ 4 for GPQA-Diamond to have sufficient sample size).
• Dataset-level power-law scaling. Treating the dataset error as the Laplace transform of the margin density p_D(m), the paper shows x^{-1/2} error scaling arises as soon as p_D(m) ∝ m^{-1/2} near 0. For the "broad" dataset families D₁ and D₂ the margin distribution satisfies lim_{m→0+} p(m) ∝ 1/√m; for the adversarial family D₃, p(m) ∝ m^{-n-1/2}, giving faster convergence.
• The theoretical dataset families look realistic. Kernel density estimates built by sampling up to 100 questions from real models and benchmarks closely resemble the D₁ and D₂ margin distributions.
• Correct answers scale fastest. Plotting true error for Llama-3.2-3B on GSM8K and Qwen-32B on GPQA-Diamond across temperatures 0.4 to 1, the paper finds very strong power-law scaling on aligned (correct-answer) questions, weaker power-law scaling on misaligned questions, consistent power-law behavior for full free-response datasets, and a less monotonic pattern for multiple-choice datasets. The authors connect this last point to Schaeffer et al. (2025), who found test-time scaling on multiple-choice tasks hard to predict, and leave a tight bound for incorrect-answer-dominated datasets as an open problem.
• Fixed allocation has a closed-form optimum. With oracle access to the margin and error model exp(−m·x), samples should be split equally among questions with equal margin. For p(m) ∝ m^{-r}, r ∈ (0,1), the optimal allocation is x_m = m^{-1}(log m − log λ) for m ≥ λ and 0 for m < λ, with λ ∼ x̄^{-1/r} as x̄ → ∞, giving error decay x̄^{-(1−r)/r}, which is x̄^{-1} when r = 1/2. Below the threshold m < λ no samples should be spent at all, because the marginal gain is smaller than adding a sample to a higher-margin question. Tests using margins extracted from Llama-3.2-3B on MATH follow the predicted allocation, with low-margin (uncertain) questions receiving more samples.
• Dynamic allocation can decay exponentially. Adapting a martingale confidence-sequence stopping rule, the paper uses the PPR-1v1 criterion (Anand Jain et al., 2022), which has theoretically optimal exponential error decay for predicting the mode. Each iteration allocates a sample to the question with lowest confidence, scored by (K−1)·Beta(x, n₁+1, n₂+1) where K is the number of unique answers and n₁, n₂ are the counts of the two most frequent answers; sampling stops when Beta(1/2, n₁+1, n₂+1) ≤ δ/(K−1) for target error δ.
• Blend-ASC wins on efficiency. Blend-ASC blends the adaptive variants and uses 4.8× fewer samples on average than vanilla SC, outperforming both fixed- and dynamic-allocation SC baselines. It is hyperparameter-free, supports batching, and can fit any budget. The paper describes it as matching the initial performance of existing SC variants in the low-sample regime while outperforming all variants at scale.
• Stated limitation. The framework assumes answers can be aggregated by a majority-vote-style rule, so Blend-ASC and its guarantees do not directly apply to tasks where aggregation is unnatural or ill-defined, such as open-ended generation.

Methodology in Plain English

The authors start from a simple reframing: self-consistency picks the most frequent answer out of many samples, which is exactly what statisticians call estimating the mode of a distribution, and what voting theorists study as a plurality election among i.i.d. voters. That reframing lets them borrow known results and derive a fresh error bound for a single question. Their bound depends on how far the top answer's probability is from the second-place answer's, measured as a "margin" that turns out to be (√p₁ − √p₂)² rather than the simpler difference p₁ − p₂. To make the bound usable when a model can produce essentially unlimited distinct answers, they lump all the rare answers into a tail, so only the top few candidates need to be tracked.

They then scale this up from one question to a whole dataset. Assuming each question's error decays exponentially in the margin, the dataset's average error becomes a Laplace transform of the distribution of margins across questions. That mathematical form is what produces the power-law curve: if small margins are distributed like m^{-1/2}, the error falls off as x^{-1/2}. The authors invent three idealized dataset families (one uniform over plausible probability pairs, one that downweights low-confidence questions, and one "adversarial" family that downweights low-margin questions aggressively) and check, using kernel density estimation on 100 real questions, that the first two resemble reality.

With the theory in hand, they ask how a fixed budget should be spent. If you knew each question's margin in advance (oracle access), the error-minimization problem is convex and has a clean closed-form solution: give more samples to uncertain questions and nothing at all to hopeless ones. Since margins are not known in advance, they also study dynamic allocation, where samples are added one at a time to whichever question currently looks least decided, and sampling stops once a confidence sequence says the leading answer is safe.

Finally they combine these ideas into Blend-ASC, which blends the adaptive strategies to fit whatever sample budget is available, and compare it against vanilla SC and fixed- and dynamic-allocation baselines. Experiments use Llama-3.2-3B, Qwen-2.5-Math and Qwen-2.5-32B on MMLU, MATH and GPQA-Diamond, plus GSM8K and GPQA-Diamond for the scaling-law plots.

Why This Matters

Self-consistency is one of the most widely deployed test-time techniques for LLM reasoning, yet its cost scales linearly with the number of samples per question, which makes dataset-wide evaluation and deployment expensive. This paper replaces the usual ad hoc statistical or semantic heuristics with a principled account of when and why SC converges, and shows how to cut sampling cost substantially. Concrete accuracy figures beyond the 4.8× sample reduction are not reported in the content available here.

Impact on research: It establishes test-time scaling laws for self-consistency, links LLM inference to classical mode estimation and voting theory, sharpens a prior per-question bound, and gives the first analysis covering fixed and dynamic allocation — while explicitly flagging multiple-choice scaling and incorrect-answer-dominated datasets as unsolved.

Real-world applications:

• Cost-efficient inference on reasoning benchmarks and production question-answering, where Blend-ASC can hit a target accuracy with roughly 4.8× fewer samples than naive self-consistency.
• Budget-constrained deployment, where the method's ability to fit any sample budget and to batch requests makes it practical when compute is capped.
• Confidence-aware serving, using the margin and stopping criteria to decide when an answer is settled and when to keep sampling, which also supports reliability-focused use cases.
• Preference optimization and data curation pipelines, since self-consistency has been used for preference optimization and for enhancing LLM reliability.

Industry relevance: Any organization serving chain-of-thought reasoning at scale pays directly for the number of samples generated. A hyperparameter-free method that cuts sample counts by a factor of 4.8, supports batching, and requires no per-task tuning is straightforward to integrate, and the theory gives engineers a principled way to reason about how much sampling is enough.

Future Directions

• Tight bounds for incorrect-answer-dominated datasets. The authors note that a tight result there would require an infeasible combinatorial analysis over the number of possible rankings between the model's most likely answer and the true answer; they leave it as an open problem.
• Multiple-choice scaling. The full-dataset plots show less monotonic scaling for multiple-choice questions, echoing Schaeffer et al. (2025), and the theory does not yet explain this regime.
• Extending beyond majority vote. Blend-ASC and its guarantees assume answers can be aggregated by a majority-vote-style rule, so open-ended generation and other tasks where aggregation is ill-defined remain out of scope.
• Reducing reliance on margin estimation. Experiments estimate margins with 100 samples per question; extending these results to settings with noisy or expensive margin estimates is a natural next step.

Target Audience

Machine learning researchers working on test-time inference, LLM reasoning, and inference efficiency; statisticians and learning theorists interested in mode estimation and voting theory applied to language models; and ML engineers or practitioners who deploy chain-of-thought reasoning at scale and need a principled, low-tuning way to cut sampling costs. The paper assumes comfort with probability bounds, exponential error rates, and optimization arguments, so a beginner audience will find the theoretical sections demanding.

Authors’ abstract

Self-consistency (SC) is a widely used test-time inference technique for improving performance in chain-of-thought reasoning. It consists of generating multiple responses, or ``samples", from a large language model (LLM) and selecting the most frequent answer. This procedure can naturally be viewed as a majority vote or empirical mode estimation. Despite its effectiveness, self-consistency is prohibitively expensive at scale when naively applied to datasets, and it lacks a unified theoretical understanding of sample efficiency and scaling behavior. In this paper, we provide the first comprehensive analysis of SC's scaling behavior and its variants, drawing on mode estimation and voting theory. We derive and empirically validate power law scaling for self-consistency across datasets, and analyze the sample efficiency for fixed-allocation and dynamic-allocation sampling schemes. From these insights, we introduce Blend-ASC, a novel variant of self-consistency that dynamically allocates samples to questions during inference, achieving state-of-the-art sample efficiency. Our approach uses 4.8 times fewer samples than vanilla SC on average, outperforming both fixed- and dynamic-allocation SC baselines, thereby demonstrating the superiority of our approach in terms of efficiency. In contrast to existing variants, we note that Blend-ASC is hyperparameter-free, supports batching, and can fit any budget of samples, ensuring it can be easily applied to any self-consistency application.