Long Live The Balance: Information Bottleneck Driven Tree-based Policy Optimization

arXiv:2605.28109 · 2026-05-28 공개 · arXiv · PDF

reinforcement-learning llm-training policy-optimization information-bottleneck online-rl monte-carlo-estimation trajectory-efficiency tree-sampling

Abstract

Recent advances in online reinforcement learning (RL) for large language models (LLMs) have demonstrated promising performance in complex reasoning tasks. However, they often exhibit an imbalanced exploration-exploitation trade-off, resulting in unstable optimization and sub-optimal performance. We introduce IB-Score, a novel metric grounded in Information Bottleneck theory that evaluates policy's exploration-exploitation balance by quantifying the trade-off between step-level reasoning diversity and mutual information shared with the correct answer. Analysis based on IB-Score shows that popular online RL approaches (e.g., GRPO) with common regularizers fail to consistently maintain balance during training with suboptimal results. To address this, we propose Information Bottleneck-driven Tree-based Policy Optimization (IB-TPO), a principled framework that formulates IB-Score as a fine-grained optimization objective and utilizes a novel IB-guided tree sampling strategy that not only improves the efficiency of online sampling with 50% more trajectories under the same token budget, but also reuses the tree structure for effective IB-Score Monte Carlo estimation. Extensive experiments across standard benchmarks show that our method significantly outperforms GRPO baseline by 2.9% to 3.6% and also outperforms other state-of-the-art online RL approaches. Our code is available at https://github.com/alibaba/EfficientRL.

한국어 요약

📋 한 줄 요약

**[LLM RL / Information Bottleneck]** IB-TPO가 IB-Score(reasoning diversity vs mutual information 균형)를 fine-grained 목표로 정식화, IB-guided tree sampling으로 같은 토큰 예산에 50% 더 많은 trajectory·tree 재사용 MC 추정, GRPO 대비 2.9~3.6% 향상.

🎯 핵심 기여도

💡 핵심 아이디어

LLM RL의 안정·강력한 학습은 reward만 최적화하는 것이 아니라 reasoning diversity와 답 정보량의 Information Bottleneck 균형을 fine-grained 목표로 두는 데서 비롯되며, tree-based sampling은 같은 예산으로 더 많은 trajectory를 만들고 tree 구조를 재사용해 IB-Score를 효율적으로 추정한다.

🔬 기술적 접근법

📊 주요 결과

💭 의의 및 한계

**의의**: LLM RL의 exploration-exploitation 균형을 IB 이론으로 원리적 정량화, GRPO의 단점(balance 유지 실패) 진단·해결, tree 구조의 sampling·추정 두 측면 동시 활용. **한계**: IB-Score 추정의 tree 구조 의존성, 50% 효율 향상의 task·모델 의존, IB-Score 외 다른 균형 metric과 비교 부재.

🚀 실용적 활용