Reproduce “Agent-Omit: Adaptive Context Omission for Efficient 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

Yansong Ning, Jun Fang, Naiqiang Tan, Hao Liu

Agent-Omit is a framework that enables Large Language Model (LLM) agents to adaptively omit redundant thoughts and observations during multi-turn interactions. This approach reduces token consumption by an average of 3 to 4 turns per trajectory while achieving comparable or superior task accuracy across five diverse benchmarks.

Key insights

  • The necessity of agent thoughts and the utility of environmental observations vary significantly across different interaction turns, with redundancy often concentrated in intermediate phases.
  • Selectively omitting specific thoughts or observations at opportune turns can reduce token cost without degrading task accuracy, demonstrating the potential for adaptive context management.
  • A combined approach of synthetic data pre-training (behavior synthesis) and reinforcement learning with a specialized reward function effectively teaches LLM agents to perform adaptive context omission.

Summary by alphaXiv.

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