Paper highlights GRPO gradient conflict flaw, proposes Bayesian fix

kastnerkyle · x · 2026-08-31

This post discusses a new arXiv paper identifying a subtle flaw in GRPO (Group Relative Policy Optimization), commonly used for post-training reasoning models. Standard GRPO assumes gradients from different queries in a mini-batch are equally reliable and averages them deterministically. However, gradients from different queries often conflict in direction, making the simple average a washed-out, inefficient update vector. The authors propose modeling group gradients as random variables within a Bayesian framework rather than fixed vectors, using a Dirichlet-based formulation to capture gradient distribution for better optimization.

Original post →

More from Research

Research channel →