Reproduce “MemArbiter: Decision-Time Memory Arbitration for Long-Horizon LLM Agents” 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
Jiajun Dong, Yutao Hu, Fengrui Fan, Shihan Dou, Yueming Wu, Deqing Zou
Researchers from Huazhong University of Science and Technology developed MemArbiter, a decision-time memory arbitration framework for long-horizon LLM agents, to address the "Memory-Action Gap" where accessible information fails to guide action selection. MemArbiter consistently improved task success rates on the ALFWorld benchmark, achieving up to 92.54% success, and enhanced recovery from failures by managing heterogeneous memories based on their functional roles and dynamic relevance.
Key insights
- The "Memory-Action Gap" highlights that simply making information accessible is insufficient; effective action guidance requires proper formation, organization, prioritization, and presentation of memories to the LLM.
- Memories in long-horizon tasks are functionally heterogeneous, demanding a function-aware management approach that differentiates their roles, persistence, and optimal activation timing.
- Dynamic adaptation of memory emphasis and presentation based on evolving task requirements and decision needs is crucial for improving agent coherence and performance over extended interactions.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.