Taming the Thinker: Conditional Entropy Shaping for Adaptive LLM Reasoning

2026-05-19

AI Summary

Existing entropy-based reasoning methods struggle to balance response length and accuracy. This work proposes a Conditional Entropy Shaping (CES) framework that dynamically regulates the conditional entropy of each token during large language model generation via DAPO reinforcement learning, adaptively adjusting reasoning depth based on problem difficulty. The key innovation lies in a conditional bidirectional policy: suppressing high-entropy branches along correct reasoning paths to enhance conciseness, while encouraging high entropy on erroneous paths to improve exploration and error correction. Experiments demonstrate that CES simultaneously improves accuracy and reduces output length across twelve mathematical benchmarks on both DeepSeek-R1-Distill-7B and 1.5B models, while maintaining robust performance on out-of-domain tasks.

Abstract

Entropy-based deep reasoning has emerged as a promising direction for improving the reasoning capabilities of Large Language Models (LLMs), but existing methods often either increase response length indiscriminately or shorten responses at the cost of accuracy. To better balance this trade-off, we introduce Conditional Entropy Shaping (CES), a framework that dynamically controls token-level response entropy, enabling LLMs to produce concise solutions on simple problems while encouraging deeper exploration on hard ones. Built on DAPO, CES uses token-level entropy as an uncertainty signal and applies a conditional bidirectional policy: it penalizes high-entropy "forking point" tokens on correct reasoning paths to improve conciseness, and rewards them on incorrect paths to encourage exploration and error correction. We implement CES on DeepSeek-R1-Distill-7B and evaluate it on 12 mathematical benchmarks. CES consistently improves average accuracy while reducing response length relative to DAPO, and supplementary experiments show similar trends on a smaller 1.5B backbone and on out-of-domain benchmarks.

Problem

• entropy-based reasoning
• response length
• accuracy trade-off
• Large Language Models
• reasoning capability

Innovation

• Conditional Entropy Shaping
• adaptive reasoning
• token-level entropy
• large language models
• DAPO