New BREW framework enables LLM agents to learn from experience

Researchers have developed BREW, a novel framework designed to enable Large Language Model (LLM)-based agents to learn from past experiences. Unlike current agents that restart learning with each session, BREW distills interaction trajectories into a structured knowledge base of natural-language recipes. This knowledge base captures essential information about task execution, applicability, and potential pitfalls. The system utilizes an Expand-and-Gather Monte Carlo Tree Search algorithm for efficient knowledge construction and retrieval, and incorporates hindsight relabeling to convert unsuccessful attempts into learning opportunities. In evaluations on benchmarks like OSWorld and tau^2-Bench, BREW demonstrated significant improvements in task success rates and reduced execution steps compared to baseline agents and existing memory-augmented systems. AI

IMPACT This framework could significantly improve the efficiency and capability of AI agents by enabling them to retain and leverage past learning.

RANK_REASON The cluster contains an academic paper detailing a new framework for LLM agents. [lever_c_demoted from research: ic=1 ai=1.0]

Read on arXiv cs.AI →

• Large Language Model
• LLM
• OSWorld
• Shashank Kirtania
• SpreadSheetBench
• tau^2-Bench

• paper
• model release
• other

AI-generated summary · Google Gemini · from 1 sources. How we write summaries →