Reproduce “LightSearcher: Efficient DeepSearch via Experiential Memory” 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

Hengzhi Lan, Yue Yu, Li Qian, Li Peng, Jie Wu, Wei Liu, Jian Luan, Ting Bai

LightSearcher is an efficient Reinforcement Learning framework for DeepSearch that resolves the accuracy-efficiency trade-off by dynamically generating and utilizing interpretable textual experiential memory and adaptive reward shaping. It achieves accuracy comparable to state-of-the-art methods while reducing search tool invocations by 39.6%, inference time by 48.6%, and token consumption by 21.2% on multi-hop QA benchmarks.

Key insights

  • Explicit, interpretable textual experiential memory can effectively guide LLMs to make more judicious DeepSearch tool-use decisions, improving both accuracy and efficiency.
  • An adaptive reward mechanism that conditionally penalizes tool usage based on answer correctness allows for significant efficiency gains without compromising the model's ability to gather necessary information.
  • Combining contrastive learning with RL training enables LLMs to learn and internalize complex strategies for balancing information gathering and computational cost, demonstrating a progressive refinement of reasoning capabilities.

Summary by alphaXiv.

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