Lambda, Stanford, and TAMU open-sourced AgentFlow, a trainable agent system utilizing Flow-GRPO turn-by-turn credit assignment to enable 7B models to beat GPT-4o on agentic and scientific benchmark suites.
Key Takeaways
- βProves specialized multi-role loops and tool integration enable 7B models to exceed frontier API accuracy;
- βFlow-GRPO assigns reinforcement learning credit turn-by-turn rather than treating trajectories as opaque units;
- βPaper, reproducible code, and training pipelines released openly for cost-effective private agent deployment.