Saved in:
Bibliographic Details
Main Authors: Chen, Hongzhan, Yang, Tao, Zhu, Yuhua, Gao, Shiping, Quan, Xiaojun, Yao, Ting
Format: Preprint
Published: 2026
Subjects:
Online Access:https://arxiv.org/abs/2603.00963
Tags: Add Tag
No Tags, Be the first to tag this record!
_version_ 1866914621256892416
author Chen, Hongzhan
Yang, Tao
Zhu, Yuhua
Gao, Shiping
Quan, Xiaojun
Yao, Ting
author_facet Chen, Hongzhan
Yang, Tao
Zhu, Yuhua
Gao, Shiping
Quan, Xiaojun
Yao, Ting
contents While reinforcement learning (RL) has been central to the recent success of large language models (LLMs), RL optimization is notoriously unstable, especially when compared to supervised fine-tuning (SFT). In this work, we investigate the stability gap between SFT and RL from a gradient-based perspective, and show that the convexity of the SFT loss with respect to model logits plays a key role in enabling stable training. Our theoretical analysis demonstrates that this property induces favorable gradient directionality during optimization. In contrast, Proximal Policy Optimization (PPO), a widely adopted policy gradient algorithm utilizing a clipped surrogate objective, lacks this stabilizing property. Motivated by this observation, we propose Logits Convex Optimization (LCO), a simple yet effective policy optimization framework that aligns the learned policy with an optimal target derived from the original RL objective, thereby emulating the stabilizing effects of logits-level convexity. Extensive experiments across multiple model families show that our LCO framework consistently improves training stability and outperforms conventional RL methods on a broad range of benchmarks.
format Preprint
id arxiv_https___arxiv_org_abs_2603_00963
institution arXiv
publishDate 2026
record_format arxiv
spellingShingle Stabilizing Policy Optimization via Logits Convexity
Chen, Hongzhan
Yang, Tao
Zhu, Yuhua
Gao, Shiping
Quan, Xiaojun
Yao, Ting
Machine Learning
Computation and Language
While reinforcement learning (RL) has been central to the recent success of large language models (LLMs), RL optimization is notoriously unstable, especially when compared to supervised fine-tuning (SFT). In this work, we investigate the stability gap between SFT and RL from a gradient-based perspective, and show that the convexity of the SFT loss with respect to model logits plays a key role in enabling stable training. Our theoretical analysis demonstrates that this property induces favorable gradient directionality during optimization. In contrast, Proximal Policy Optimization (PPO), a widely adopted policy gradient algorithm utilizing a clipped surrogate objective, lacks this stabilizing property. Motivated by this observation, we propose Logits Convex Optimization (LCO), a simple yet effective policy optimization framework that aligns the learned policy with an optimal target derived from the original RL objective, thereby emulating the stabilizing effects of logits-level convexity. Extensive experiments across multiple model families show that our LCO framework consistently improves training stability and outperforms conventional RL methods on a broad range of benchmarks.
title Stabilizing Policy Optimization via Logits Convexity
topic Machine Learning
Computation and Language
url https://arxiv.org/abs/2603.00963