Microsoft Proposes GFlowRL for Scaling Reasoning Models

dair_ai · x · 2026-07-17

A new paper from Microsoft discusses scaling GFlowNet-style distribution-matching reinforcement learning to large-scale reasoning models.

The authors point out that GFlowNet methods don't force the reward onto a single optimal path; instead, they let the model learn a "reward distribution," encouraging diverse reasoning trajectories. However, as models grow larger, rollouts get longer, and reward noise increases, the prompt-conditional partition function that needs to be learned introduces gradient instability and extra engineering complexity.

The proposed GFlowRL eliminates this learned partition function entirely, replacing it with an in-batch Monte Carlo estimation using rollout groups generated during training. The authors claim this reduces engineering overhead and makes the method better suited for large-scale reasoning model training.

Original post →

More from Research

Research channel →