Ppo critics suffer from value flattening that sparse updates reduce
Rethinking Critic Learning in PPO: Understanding and Mitigating Value Flattening
Machine LearningArtificial Intelligence
Summary
In reinforcement learning, a method called Proximal Policy Optimization (PPO) uses a value-predicting part, called the critic, to help improve decisions. The authors discovered that critics often get stuck giving nearly the same value estimates even when actual state values vary a lot, a problem they call Value Flattening. They showed this effect worsens with larger problems and found it is caused by how the learning updates are done. To fix this, they propose checking and updating only a few well-chosen states at a time, which helps the critic learn more accurate values and leads to better overall policies.
What this means in practice
- •For ai training engineers: Improve policy learning stability and accuracy in training large language models by using sparse critic update methods to avoid value flattening.
- •For gaming ai developers: Enhance reinforcement learning agents in complex game environments by reducing value estimation errors through sparse critic supervision.
Authors
Yizhuo Li, Jianhao Yan, Yun Luo, Zhi Wang, Futing Wang, Rong-Xi Tan, Kanghui Tian, Ganqu Cui, Ning Ding, Peilin Zhao, Yafu Li, Yu Cheng
Abstract
In reinforcement learning for large language models, Proximal Policy Optimization (PPO) commonly uses a critic to estimate state values and reduce the variance of policy updates. However, we uncover a systematic failure mode in PPO critics, which we call Value Flattening: state values, estimated from multiple Monte Carlo continuations, change sharply across intermediate states while critic predictions remain comparatively flat. We further observe this phenomenon in a controlled FrozenLake environment and find that it becomes more pronounced as the state space grows. Our theoretical and empirical analyses relate Value Flattening to an implicit variance penalty in the critic loss and redundant updates from temporally correlated states with similar gradients. Motivated by these findings, we introduce SParse Proximal Policy Optimization (SP$^3$O), which applies the value loss to only a few well-separated states in each response to mitigate both effects. Experiments on Qwen3-Base show that SP$^3$O with only three states supervised per response can mitigate Value Flattening and consistently improve the learned policy across model sizes and evaluation suites. Together, our results identify Value Flattening as an important yet overlooked failure mode of critic learning in standard PPO and show that a simple sparse supervision strategy can mitigate it.