Reproduce “Joint Optimization of Multi-agent Memory System” 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

Wenyu Mao, Haoyang Liu, Haosong Tan, Yaorui Shi, Jiancan Wu, An Zhang, Xiang Wang

A framework named CoMAM jointly optimizes multi-agent memory systems for Large Language Models, enhancing their performance in long-horizon conversational tasks by explicitly modeling inter-agent dependencies and adaptively assigning credit. This approach achieved query-answering accuracy improvements up to +16.7% over previous state-of-the-art RL-based baselines on benchmarks like PersonaMem.

Key insights

  • Modeling inter-agent dependencies as a sequential MDP allows for explicit co-adaptation among agents during end-to-end training, leading to more cohesive system behavior.
  • An adaptive credit assignment strategy provides targeted feedback to specialized agents, balancing local task specialization with contributions to the overall global objective.
  • Joint optimization with differentiated credit assignment resolves issues of sub-optimal collaboration and ambiguous credit, fostering more effective and specialized agent roles.

Summary by alphaXiv.

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