Reproduce “EPO: Hierarchical LLM Agents with Environment Preference Optimization” on OpenResearch

Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.

Universal app for Apple silicon and Intel. Requires macOS 11 or later.

About this paper

Qi Zhao, Haotian Fu, Chen Sun, George Konidaris

The EPO framework from Brown University researchers presents a hierarchical Large Language Model (LLM) agent that learns long-horizon tasks by leveraging multimodal environmental feedback and a novel Environment Preference Optimization. This method achieved first place on the challenging ALFRED public leaderboard with a 12.0% improvement in unseen success rate compared to previous methods, demonstrating efficient learning from unannotated data.

Key insights

  • Hierarchical decomposition allows LLMs to effectively manage long-horizon tasks by breaking them into shorter, more manageable subgoals and action sequences.
  • Automatic reward modeling from multimodal environment feedback enables agents to learn efficiently from unannotated data, reducing reliance on extensive human labeling.
  • Extending preference optimization with a token-level alignment loss in EPO improves training stability and ensures the generation of desired output formats for LLM agents.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.