Reproduce “Multi-Turn Agentic Scientific Literature Search via Workflow Induction” 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
Jisen Li, Bingxuan Li, Nanyi Jiang, Xuying Ning, Xiyao Wang, Yifan Shen, Heng Wang, Yuqing Jian and 5 more
The PAPERPILOT system introduces a multi-turn agentic scientific literature search framework that utilizes explicit, editable workflow induction. It enhances retrieval quality and consistency by enabling large language models to dynamically construct and refine search strategies based on evolving user feedback. The approach, even with a 9B parameter model, achieved competitive retrieval performance and demonstrated superior cost-effectiveness compared to larger proprietary systems.
Key insights
- Adaptive workflow induction significantly improves retrieval quality by enabling query-specific search strategies, outperforming fixed pipelines across various metrics for strong tool-using agents.
- Multi-turn interaction effectively enhances search performance when user feedback is directly translated into concrete modifications of the underlying explicit search workflow.
- A smaller, purpose-trained language model (PAPERPILOT-9B) can achieve high retrieval quality and stability in workflow execution, demonstrating a favorable cost-performance trade-off compared to larger, general-purpose proprietary systems.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.