Clipping-Free Policy Optimization for Large Language Models

TL;DR

This paper introduces Clipping-Free Policy Optimization (CFPO), a novel reinforcement learning method for large language models that avoids clipping issues, improves training stability, and maintains performance across reasoning and alignment tasks.

Contribution

CFPO replaces heuristic clipping with a convex quadratic penalty based on Total Variation divergence, providing a stable, differentiable objective with minimal code changes.

Findings

CFPO matches clipping-based methods on downstream benchmarks.

CFPO extends stable training regimes and mitigates verbosity exploitation.

CFPO reduces capability degradation while maintaining instruction-following performance.

Abstract

Reinforcement learning has become central to post-training large language models, yet dominant algorithms rely on clipping mechanisms that introduce optimization issues at scale, including zero-gradient regions, reward hacking, and training instability. We propose Clipping-Free Policy Optimization (CFPO), which replaces heuristic clipping with a convex quadratic penalty derived from Total Variation divergence constraints, yielding an everywhere-differentiable objective that enforces stable policy updates without hard boundaries. We evaluate CFPO across both reasoning and alignment settings. In reasoning, CFPO matches clipping-based methods on downstream benchmarks while extending the stable training regime. In alignment, CFPO mitigates verbosity exploitation and reduces capability degradation, while achieving competitive instruction-following performance. CFPO requires only a one-line code change and no additional hyperparameters. Our results suggest that CFPO is a promising drop-in alternative to clipping-based methods for LLM post-training.