Decision-Making Models in Long-Horizon Agents: Actor, Judge, or Teacher?

Written by

in

A study compared the performance of three decision-making models and three general-purpose LLMs as actors and judges in long-horizon agents, and evaluated JEV’s effectiveness as a teacher for guiding reinforcement learning policies. On ALFWorld and WebShop, JEV increased the success rates of GRPO-based Qwen2.5-1.5B policies from 72.8% to 82.9% and from 56.8% to 71.4%, respectively.

Source:Hugging Face 社区论文精选 · Read Original

Comments

Leave a Reply

Your email address will not be published. Required fields are marked *