DVPO: Distributional Value Modeling-based Policy Optimization for LLM Post-Training
This addresses the challenge of deploying LLMs in real-world scenarios with unreliable supervision, offering a method to improve stability and generalization, though it appears incremental as it builds on existing RL approaches.
The paper tackles the problem of noisy or incomplete supervision in reinforcement learning for LLM post-training, which can destabilize training and harm generalization, and introduces DVPO, a framework that combines conditional risk theory with distributional value modeling to balance robustness and generalization, outperforming methods like PPO and GRPO in experiments on multi-turn dialogue, math reasoning, and scientific QA.
Reinforcement learning (RL) has shown strong performance in LLM post-training, but real-world deployment often involves noisy or incomplete supervision. In such settings, complex and unreliable supervision signals can destabilize training and harm generalization. While existing approaches such as worst-case optimization (e.g., RFQI, CQL) and mean-based methods (e.g., PPO, GRPO) can improve stability, they often overlook generalization and may produce overly conservative policies, leading to uneven performance across diverse real scenarios. To this end, we introduce DVPO (Distributional Value Modeling with Risk-aware Policy Optimization), a new RL framework that combines conditional risk theory with distributional value modeling to better balance robustness and generalization. DVPO learns token-level value distributions to provide fine-grained supervision, and applies an asymmetric risk regularization to shape the distribution tails: it contracts the lower tail to dampen noisy negative deviations, while expanding the upper tail to preserve exploratory diversity. Across extensive experiments and analysis in multi-turn dialogue, math reasoning, and scientific QA, DVPO consistently outperforms PPO, GRPO, and robust Bellman-based PPO under noisy supervision, showing its potential for LLM post-training in the real-world.