PaSa - ByteDance's intelligent agent for searching academic papers

What is PaSa?

PaSa is an AI agent for academic paper retrieval developed by ByteDance Research, based on reinforcement learning. It mimics the behavior of human researchers, automatically calling search engines, browsing relevant papers, and tracking citation networks to provide users with accurate and comprehensive academic paper search results.

The main functions of PaSa

• Automatically call search toolsPaSa can automatically call a search engine, generate diverse search keywords based on the academic questions entered by the user, and perform multiple searches to ensure comprehensive coverage of relevant literature.
• Reading and analyzing the content of the paperPaSa achieves efficient information processing through its core components, Crawler and Selector. Crawler is responsible for collecting relevant papers, including discovering more potentially relevant literature by expanding the citation network. Selector then carefully reads the collected papers, filtering out those that truly meet the user's needs.
• Select relevant referencesPaSa can filter the most relevant references from a massive amount of literature, providing comprehensive and accurate search results.
• Supports complex academic queriesPaSa is designed to handle complex academic problems and can understand and process fine-grained academic queries, such as those involving specific algorithms or research methods.
• Reinforcement learning optimizationPaSa uses reinforcement learning methods for training, and improves search efficiency and accuracy by using the synthetic dataset AutoScholarQuery and the real-world query benchmark RealScholarQuery.
• High-efficiency search capabilityPaSa boasts extremely fast search speeds, capable of completing a comprehensive academic survey within two minutes. In performance tests, PaSa outperformed Google Scholar by 37.78% and 39.90% in Recall@20 and Recall@50 metrics, respectively.

PaSa's technical principles

• Core componentsPaSa's core consists of two main LLM agents: Crawler and Selector.
  • CrawlerResponsible for collecting academic papers relevant to user queries through search engines. Capable of generating diverse search keywords and executing multiple searches to maximize recall.
  • Selector: Responsible for carefully reading every paper found by Crawler, evaluating whether it meets user needs, and improving the accuracy of search results.
• Reinforcement learning optimizationPaSa is trained using reinforcement learning (RL) and proximal policy optimization (PPO) algorithms to address the sparse reward and long-trajectory problems in literature search. During training, PaSa improves performance through the following methods:
  • Synthetic datasetThe research team constructed the AutoScholarQuery dataset, which contains 35,000 fine-grained academic queries and their corresponding papers. The data comes from papers published at top artificial intelligence conferences.
  • Real-world datasetsTo evaluate PaSa's performance in real-world scenarios, the team also developed the RealScholarQuery dataset, which collects real-world academic queries.

PaSa's project address

• Project official website:pasa-agent.ai
• GitHub repository:https://github.com/bytedance/pasa
• arXiv technical paper:https://arxiv.org/pdf/2501.10120

PaSa application scenarios

• Literature review in academic researchPaSa helps researchers quickly search for and conduct research on academic papers. By mimicking the behavior of human researchers, it autonomously uses search engines, reads papers, and tracks citation networks, enabling a comprehensive academic survey to be completed in just two minutes.
• Support for scientific research and teaching in universitiesIn university settings, PaSa can serve as a research support tool, helping faculty and students quickly access relevant academic resources. Faculty can use PaSa to rapidly retrieve content from academic resource databases, assisting in tasks such as generating literature reviews, designing research experiments, and translating and polishing papers.
• Intellectual Property AnalysisPaSa's efficient search capabilities can also be applied to the field of intellectual property.
• Multi-task learning and data miningResearch from the PASA Big Data Lab at Nanjing University shows that PaSa's underlying technology can also be applied to multi-task learning and data mining.