PaSa: An LLM Agent for Comprehensive Academic Paper Search

Yichen He, Guanhua Huang, Peiyuan Feng, Yuan Lin, Yuchen Zhang, Hang Li, Weinan E

Abstract

Conducting a thorough literature review is a cornerstone of scientific research, yet existing academic search engines often struggle with complex, fine-grained queries. We present PaSa (Paper Search), an autonomous LLM agent designed to mimic the iterative research habits of human experts—searching, reading, and following citation trails. Built on a dual-agent architecture and trained via reinforcement learning on a novel synthetic dataset, PaSa-7B significantly outperforms traditional search engines and frontier models like GPT-4o in retrieving highly relevant papers for specific, long-tail academic requests.

Introduction

While standard academic search engines like Google Scholar are effective for broad queries, they often fail to address highly specific research questions, such as identifying studies that focus on non-stationary reinforcement learning using specific UCB-based algorithms. Human researchers solve this by performing iterative "tree-searches": they read abstracts, examine related work sections, and "hop" through citation networks to find relevant literature.

Existing LLM-based tools typically attempt to solve this by simply rewriting search queries, failing to replicate the deep, iterative navigation required for a comprehensive review. PaSa addresses this gap by creating an agentic system that autonomously navigates the web and citation networks just as a human researcher would.

System Architecture

PaSa is built on a two-agent architecture utilizing the Qwen2.5-7B model as its backbone. The system splits the labor of discovery and evaluation into two distinct roles:

  1. The Crawler: This agent is responsible for exploring the literature landscape. It can execute three primary actions: [Search] (generating queries for arXiv or Google), [Expand] (reading a specific subsection of a paper to extract citations into a queue), and [Stop] (concluding the search for a specific paper and moving to the next candidate).
  2. The Selector: This agent acts as a high-precision filter. It reviews the title and abstract of every paper identified by the Crawler, outputting a "True/False" relevance decision accompanied by a rationale explaining how the paper meets the user's specific criteria.

Training via Reinforcement Learning

Training the Crawler agent presents significant challenges due to sparse rewards and long trajectory lengths; a single search session may involve hundreds of papers before a relevant one is found. To overcome this, the authors utilized the AGILE reinforcement learning framework and a novel session-level PPO (Proximal Policy Optimization) algorithm.

To address the sparse reward problem, the Selector agent was utilized as an auxiliary reward model. This provided the Crawler with denser feedback during training, allowing it to learn which citation paths and search queries were most likely to lead to relevant results.

Datasets and Benchmarks

The project introduced two critical datasets to facilitate training and evaluation:

  • AutoScholarQuery: A synthetic training dataset of ~35,000 query-paper pairs. This was created by reverse-engineering the "Related Work" sections of top-tier AI conference papers (ICLR, NeurIPS, CVPR). By feeding these sections into GPT-4o, the authors generated the specific queries that those citations were originally intended to answer.
  • RealScholarQuery: A high-quality benchmark consisting of 50 complex academic queries. These were manually curated and annotated by computer science professors to serve as a rigorous, real-world evaluation of the agent's performance.

Key Results and Insights

Evaluation on the RealScholarQuery benchmark demonstrated that PaSa-7B provides a substantial leap in retrieval performance:

  • Superior Recall: PaSa-7B outperformed the best traditional baseline (Google paired with GPT-4o query paraphrasing) by 37.78% in Recall@20 and 39.90% in Recall@50.
  • Efficiency vs. Scale: Surprisingly, the 7B-parameter model trained with task-specific RL outperformed a version of the same agent powered by GPT-4o (PaSa-GPT-4o) by 30.36% in recall. This suggests that specialized RL training is more effective for navigation tasks than general-purpose scaling.
  • High-Precision Filtering: The 7B Selector agent achieved an F1 score of 85%, surpassing GPT-4o’s 80% on the same relevance-testing task.
  • Ensemble Performance: By running the Crawler multiple times using sampling decoding, the system reached a final recall of nearly 65% on complex, real-world queries.

Conclusion and Future Work

PaSa demonstrates that agentic workflows—specifically those that utilize citation crawling and iterative reading—are significantly more effective for academic search than single-pass retrieval systems. While currently focused on the field of Artificial Intelligence, future iterations of PaSa aim to expand into other scientific domains such as biology and physics, and scale the underlying models to further enhance reasoning and navigation capabilities.

Summary updated

BibTeX
@inproceedings{he2025pasa,
  author = {Yichen He and Guanhua Huang and Peiyuan Feng and Yuan Lin and Yuchen Zhang and Hang Li and Weinan E},
  title = {{PaSa: An LLM Agent for Comprehensive Academic Paper Search}},
  booktitle = {Annual Meeting of the Association for Computational Linguistics},
  year = {2025},
}
Figure