Think Just Enough: Sequence-Level Entropy as a Confidence Signal for LLM Reasoning

2025-10-09

AI Summary

To address high computational overhead and low token efficiency in large language model (LLM) inference, this paper proposes an entropy-based early-exit framework. It leverages the Shannon entropy of token-level log-probabilities in the model’s output sequence as a confidence signal to dynamically assess answer convergence and terminate generation early. The key contribution is the discovery—first reported here—that advanced reasoning models exhibit an emergent “entropy-aware correctness” capability: correct answers consistently exhibit lower sequence entropy, enabling entropy thresholds to serve as a cross-model generalizable confidence metric. This metric requires only minimal calibration (few-shot) and imposes no architectural constraints, ensuring compatibility with diverse reasoning-optimized models. Experiments across multiple reasoning benchmarks demonstrate 25–50% reduction in computational cost—measured in FLOPs and latency—without accuracy degradation, while significantly improving inference latency stability and energy efficiency.

Technology Category

• Knowledge Representation and Reasoning: Computational Complexity of Reasoning
• Cognitive Modeling & Cognitive Systems: Conceptual Inference and Reasoning
• Machine Learning: Large Multimodal Models (LMMs)

Application Category

• Economics, Online Markets and Human Computation: Cost models of using LLMs in production systems
• Graph Algorithms and Modeling for the Web: Foundation models and LLMs for Web-related graphs
• Semantics and Knowledge: Data modeling to support human-machine intelligence, including LLMs agents, intelligent system behavior, explanations, and user-friendly interactions

Abstract

We introduce a simple, yet novel entropy-based framework to drive token efficiency in large language models during reasoning tasks. Our approach uses Shannon entropy from token-level logprobs as a confidence signal to enable early stopping, achieving 25-50% computational savings while maintaining task accuracy. Crucially, we demonstrate that entropy-based confidence calibration represents an emergent property of advanced post-training optimization present in modern reasoning models but notably absent in standard instruction-tuned and pre-trained models (Llama 3.3 70B). We show that the entropy threshold to stop reasoning varies from model to model but can be calculated easily in one shot using only a few examples from existing reasoning datasets. Our results indicate that advanced reasoning models often know that they've gotten a correct answer early on, and that this emergent confidence awareness can be exploited to save tokens and reduce latency. The framework demonstrates consistent performance across reasoning-optimized model families with 25-50% computational cost reduction while preserving accuracy, revealing that confidence mechanisms represent a distinguishing characteristic of modern post-trained reasoning systems versus their predecessors.

Problem

• Reducing computational costs in LLM reasoning tasks via entropy-based early stopping
• Identifying emergent confidence calibration in post-trained reasoning models
• Enabling token-efficient reasoning while maintaining task accuracy performance

Innovation

• Uses entropy-based confidence for early stopping
• Achieves 25-50% computational savings with accuracy
• Requires only few examples to calculate entropy threshold