Hugging Face Training Agents 3 explains GRPO for outcome-based agent training

Hugging Face · youtube · 2026-07-22

Hugging Face’s Training Agents 3 session covers reinforcement learning with GRPO, moving from a fine-tuned baseline to outcome-based training.

The session explains why GRPO comes after SFT and what RL adds beyond imitation. It walks through group sampling, relative advantages, and how GRPO replaces a critic or reward model by using the group’s relative scores as the learning signal. The agenda also includes writing verifiable reward functions for agent tasks, running GRPO in TRL on Hugging Face Jobs, and reading training curves such as reward, KL, and completion length.

The session also calls out common failure modes: reward hacking, collapse, and length gaming. It is part of a broader series about using coding agents to design, run, monitor, and review post-training experiments for better agentic models.

Related event: Hugging Face Live Stream Covers GRPO for Agent Reinforcement Learning(2 posts)→

Original post →

More from coding & agent

coding & agent channel →