Smooth Gate Functions for Soft Advantage Policy Optimization

TL;DR

This paper explores the use of smooth gate functions in policy optimization for large language models, demonstrating improved stability and performance over traditional hard clipping methods.

Contribution

It introduces a formal framework for admissible gate functions, evaluates various smooth gates, and provides empirical insights for enhancing policy optimization stability.

Findings

Smooth sigmoid-based gates improve training stability.

Different gate functions impact final model performance.

Guidelines for designing robust policy optimization objectives.

Abstract

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.