Reinforcement learning (RL) has emerged as a core approach for enhancing the reasoning abilities of large language models (LLMs). However, existing RL methods often suffer from premature entropy collapse, selective over-updating of high-reward traject...
Reinforcement learning (RL) has emerged as a core approach for enhancing the reasoning abilities of large language models (LLMs). However, existing RL methods often suffer from premature entropy collapse, selective over-updating of high-reward trajectories, and instability arising from poorly controlled optimization dynamics. Crucially, current approaches lack a principled mechanism for constraining policy updates at the prompt level, which leads to inconsistent learning behavior across trajectories. We introduce Prompt-wise Trust-Region Optimization (PTRO), a new algorithm that formulates RL finetuning as an exact prompt-wise trust-region optimization problem. This formulation yields a principled dual update that explicitly regulates per-prompt update magnitude and removes the need for heuristic clipping or global KL penalties. PTRO achieves higher Pass@k, slower entropy decay, and significantly more stable training, remaining robust even under larger learning rates and deeper inner-loop updates. Our results demonstrate that prompt-wise trust-region optimization provides a simple, principled, and highly effective foundation for stable RL fine-tuning of LLMs.