arXivDaily arXiv每日学术速递 周一至周五更新
arXiv周末暂无论文更新,休息一下吧,周末愉快~~
arXiv 2604.26326cs.LGcs.CLstat.ML
收藏

通过精确熵曲线控制解决LLM RL中的性能饱和问题

Addressing Performance Saturation for LLM RL via Precise Entropy Curve Control

Purdue University(普渡大学)

Bolian Li, Yifan Wang, Yi Ding, Anamika Lochab, Ananth Grama, Ruqi Zhang

AI总结:

本文提出Entrocraft方法,通过偏差优势分布实现用户定制的熵调度,解决LLM RL中的性能饱和问题,提升泛化能力、输出多样性及长期训练效果。

摘要:

强化学习(RL)使大语言模型(LLMs)具备复杂推理能力。然而,大多数RL算法面临性能饱和问题,随着训练规模扩大,无法持续提升。该问题可通过熵崩溃体现,是RL探索的关键诊断指标。现有方法通过正则化或裁剪防止熵崩溃,但其熵曲线在长期中不稳定,阻碍性能提升。本文提出Entrocraft,一种简单的拒绝采样方法,通过偏置优势分布实现用户定制的熵调度。Entrocraft无需目标正则化,且对优势估计器无关。理论上,我们将在最小假设下将每步熵变化与优势分布相关联。这解释了现有RL和熵保持方法的行为。Entrocraft还使熵调度的系统研究成为可能,揭示线性退火(从高开始衰减到略低的目标)表现最佳。实验证明,Entrocraft解决性能饱和问题,显著提升泛化能力、输出多样性和长期训练效果。它使4B模型超越8B基线,持续改进时间延长至4倍,且将pass@K提升50%。

英文摘要:

Reinforcement learning (RL) has enabled complex reasoning abilities in large language models (LLMs). However, most RL algorithms suffer from performance saturation, preventing continued gains as RL training scales. This problem can be characterized by the collapse of entropy, a key diagnostic for exploration in RL. Existing attempts focus on preventing entropy collapse through regularization or clipping. However, their resulting entropy curves often exhibit instability in the long term, which hinders performance gains. In this paper, we introduce Entrocraft, a simple rejection-sampling approach that realizes user-customized entropy schedule by biasing the advantage distributions. Entrocraft requires no objective regularization and is advantage-estimator-agnostic. Theoretically, we relate per-step entropy change to the advantage distribution under minimal assumptions. This explains the behavior of existing RL and entropy-preserving methods. Entrocraft also enables a systematic study of entropy schedules, which reveals that linear annealing, which starts high and decays to a slightly lower target, performs best. Empirically, Entrocraft addresses performance saturation, significantly improving generalization, output diversity, and long-term training. It enables a 4B model to outperform an 8B baseline, sustains improvement for up to 4x longer before plateauing, and raises pass@K by 50% over the baseline.

机构、总结、中文摘要均由AI分析而来,可能存在错误,请以论文正文为准。

↑