Smooth Gate Functions for Soft Advantage Policy Optimization
This work addresses training instability for large language model developers, but it is incremental as it builds on existing soft policy optimization methods.
The paper tackled the instability in Group Relative Policy Optimization (GRPO) for large language models by investigating different smooth gate functions, finding that replacing hard clipping with these functions improves training stability and performance, with experiments on Qwen2.5-7B-Instruct showing concrete gains in mathematical reasoning tasks.
Group Relative Policy Optimization (GRPO) has significantly advanced the training of large language models and enhanced their reasoning capabilities, while it remains susceptible to instability due to the use of hard clipping. Soft Adaptive Policy Optimization (SAPO) addresses this limitation by replacing clipping with a smooth sigmoid-based gate function, which leads to more stable updates. We have decided to push this theory further and investigate the impact of different gate functions on both training stability and final model performance. We formalize the key properties that admissible gates should satisfy and identify several families of such functions for empirical evaluation. This paper presents an analysis of our findings based on experiments conducted with the Qwen2.5-7B-Instruct model on mathematical reasoning tasks. These results provide practical guidance for designing smoother and more robust policy optimization objectives for large language model training.