Skip to main content Skip to navigation

PaSa: An LLM Agent for Comprehensive Academic Paper Search

Project Overview

The document highlights the innovative use of generative AI in education through the development of PaSa, an advanced paper search agent that leverages large language models (LLMs) to enhance the efficiency and accuracy of academic paper searches. By autonomously managing search queries, reading scholarly papers, and navigating citation networks, PaSa delivers comprehensive results for intricate academic inquiries. Its optimization is driven by reinforcement learning, and it is trained on a specialized dataset known as AutoScholarQuery, which encompasses fine-grained academic queries. The findings indicate that PaSa outperforms traditional academic search tools such as Google Scholar and ChatGPT, demonstrating superior recall and precision metrics. Overall, the document underscores the potential of generative AI to revolutionize research methodologies in education, making academic resources more accessible and effective for scholars and students alike.

Key Applications

PaSa - an LLM agent for comprehensive academic paper search

Context: Research in AI and related fields, targeting academic researchers and students

Implementation: PaSa employs two LLM agents, Crawler and Selector, optimized using reinforcement learning within the AGILE framework. It utilizes synthetic datasets for training and evaluation.

Outcomes: PaSa outperforms traditional academic search engines, achieving improved recall and precision metrics in both synthetic and real-world scenarios.

Challenges: Challenges include the complexity of queries, the need for high-quality training data, and computational resource constraints.

Implementation Barriers

Data Quality

The need for high-quality academic search data for effective training.

Proposed Solutions: Creation of synthetic datasets like AutoScholarQuery and RealScholarQuery to enhance training quality.

Resource Constraints

High computational resources required for training large models.

Proposed Solutions: Utilizing smaller models initially and planning for future expansions with larger models to enhance performance.

Project Team

Yichen He

Researcher

Guanhua Huang

Researcher

Peiyuan Feng

Researcher

Yuan Lin

Researcher

Yuchen Zhang

Researcher

Hang Li

Researcher

Weinan E

Researcher

Contact Information

For information about the paper, please contact the authors.

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

Source Publication: View Original PaperLink opens in a new window

Project Contact: Dr. Jianhua Yang

LLM Model Version: gpt-4o-mini-2024-07-18

Analysis Provider: Openai

Let us know you agree to cookies