Learning When to Sample: Confidence-Aware Self-Consistency for Efficient LLM Chain-of-Thought Reasoning

TL;DR

This paper presents a confidence-aware framework that adaptively chooses between single and multiple reasoning paths in LLMs, significantly reducing inference costs while maintaining high accuracy in reasoning tasks.

Contribution

It introduces a novel decision framework that analyzes a single reasoning trajectory to efficiently balance accuracy and computational cost without additional fine-tuning.

Findings

Achieves comparable accuracy to multi-path methods with up to 80% fewer tokens.

Effectively generalizes across multiple datasets without fine-tuning.

Utilizes signals from reasoning trajectories for uncertainty estimation.

Abstract

Large language models (LLMs) achieve strong reasoning performance through chain-of-thought (CoT) reasoning, yet often generate unnecessarily long reasoning paths that incur high inference cost. Recent self-consistency-based approaches further improve accuracy but require sampling and aggregating multiple reasoning trajectories, leading to substantial additional computational overhead. This paper introduces a confidence-aware decision framework that analyzes a single completed reasoning trajectory to adaptively select between single-path and multi-path reasoning. The framework is trained using sentence-level numeric and linguistic features extracted from intermediate reasoning states in the MedQA dataset and generalizes effectively to MathQA, MedMCQA, and MMLU without additional fine-tuning. Experimental results show that the proposed method maintains accuracy comparable to multi-path baselines while using up to 80\% fewer tokens. These findings demonstrate that reasoning trajectories contain rich signals for uncertainty estimation, enabling a simple, transferable mechanism to balance accuracy and efficiency in LLM reasoning.