Reproduce “Risk-Sensitive RL for Alleviating Exploration Dilemmas in Large Language Models” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Yuhua Jiang, Jiawei Huang, Yufeng Yuan, Xin Mao, Yu Yue, Qianchuan Zhao, Lin Yan
RS-GRPO changes the advantage calculation in GRPO to weight higher-reward responses more strongly, with β controlling the tilt. For Qwen2.5-Math-1.5B trained on DeepMath103K, it reports AIME24 Pass@32 of 45.1% versus 36.4% for GRPO, while Pass@1 is 16.7% versus 16.6%. These results do not hold against base models at high k for every model.
Key insights
- A risk-seeking objective function, controlled by the hyperparameter "beta", effectively encourages LLMs to explore low-probability regions of the solution space and discover novel reasoning paths.
- The risk-sensitive advantage function dynamically amplifies learning signals for high-reward samples, particularly correct responses on challenging prompts, fostering deeper exploration.
- Theoretical analysis demonstrates that, for a sufficiently large risk-sensitivity parameter, this approach can guarantee an increase in the probability of optimal actions, enabling escape from local optima.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.