Fixing PPO Training Noise: Value Estimation as Classification Task

joecole · x · 2026-08-08

In reinforcement learning with binary rewards, standard PPO typically trains the critic using MSE regression. However, minor value errors can significantly amplify noise in advantage functions.

A recent paper proposes a simple fix: stop regressing scalar values and treat value estimation as a classification task. The authors replace the MSE head with a categorical predictor over value bins (HL-Gauss PPO), trained using cross-entropy, and then decode the distribution back to a scalar for standard GAE.

Tested on Qwen models across math and reasoning tasks, this method consistently outperforms standard PPO and DAPO, yielding lower-variance advantages and better calibration.

Original post →

More from Research

Research channel →