• Agentic RL extends LLM training beyond single-turn responses to full decision-making via environment interaction. • It collects on‑policy data, optimizing policies across multi‑step trajectories in simulated or real settings. • Credit is assigned over long horizons, linking intermediate actions like tool selection to final outcomes. • Iterative closed‑loop training uses rollouts, reward computation, policy updates, and fresh data collection (GRPO/PPO). • GPT‑OSS demonstrates comparable performance to OpenAI’s o3‑mini while enabling agentic capabilities. • Agentic RL builds scalable, reliable AI agents that reason, adapt, and execute multi‑step workflows for professionals.

Article Summaries:

  • Summary

Researchers report progress in training GPT‑OSS, an open‑source large language model, for agentic reinforcement learning (RL). Unlike single‑turn RL, agentic RL optimizes multi‑step decision policies through active interaction with environments, assigning credit across long horizons. The team used the verl framework and benchmark tasks (gsm8k, Retool, verifiable instruction following) to fine‑tune GPT‑OSS‑20B, with similar fixes applicable to GPT‑OSS‑120B. They also benchmarked Qwen‑2.5‑32B to track metric trends. Key challenges included adapting verl to GPT‑OSS’s new Harmony chat template and ensuring correct message parsing and tool invocation. The work demonstrates a viable path to end‑to‑end agentic training for GPT‑OSS.

Sources: