PaperScout
An autonomous LLM-based agent that reformulates academic paper search as a multi-turn decision-making process, dynamically deciding when and how to invoke search and citation expansion tools.
Project Repo →Our Vision
The 5-Level Cognitive Hierarchy
AI's capability in scientific literature mining follows a progressive path of cognitive depth.
L1
Paper Search
Retrieval from massive literature pools (e.g., PaperScout).
→
L2
Element Interpretation
Parsing PDF into structured, machine-readable data (e.g., ChemTable).
→
L3
Information Extraction
Understanding methods, tasks, and contributions (e.g., ScholarSum).
→
L4
Knowledge Comprehension
Reasoning and multi-paper synthesis (e.g., PaperArena, Mind2Report).
→
L5
Hypothesis Discovery
Autonomous discovery and research generation (Future Vision).
Research Projects
Advancing the Frontiers of Science
Our work spans the entire hierarchy, providing tools and benchmarks for the next generation of scientific AI.
ChemTable
A large-scale benchmark for multimodal LLMs to recognize and understand complex chemical tables, combining symbolic chemical formulas, table structures, and visual molecule diagrams.
Project Repo →ScholarSum
Advancing scientific summarization through knowledge graph reasoning and reflective refinement, distilling complex research into structured, high-fidelity summaries.
OpenReview →PaperArena
An evaluation benchmark for tool-augmented agentic reasoning on scientific literature, requiring agents to integrate information across multiple papers with diverse tools.
Project Repo →Mind2Report
A cognitive deep research agent that emulates commercial analysts to synthesize expert-level reports from massive web sources through intent-driven search and iterative synthesis.
Project Repo →AI Scientist Vision
Our ultimate goal: building agents that not only understand literature but also autonomously discover laws, generate new hypotheses, and propose novel research directions.
Visionary ResearchTimeline
Milestones in Scientific Literature Mining
Jun. 2025
ChemTable · Table Recognition Benchmark
A large-scale benchmark designed to test MLLMs on real-world chemical tables, addressing structural complexity and domain-specific semantics.
Oct. 2025
ScholarSum · Student-Teacher Summarization
A novel Student-Teacher framework for scientific summarization, leveraging knowledge graph reasoning and reflective refinement to ensure structural coherence.
Oct. 2025
PaperArena · Agentic Reasoning Benchmark
The first benchmark for tool-augmented agentic reasoning on scientific literature, evaluating cross-paper integration and multi-tool orchestration.
Jan. 2026
PaperScout · Autonomous Paper Search
An adaptive agentic framework for academic paper search, optimized with Proximal Sequence Policy Optimization (PSPO) for multi-turn interactions.
Jan. 2026
Mind2Report · Cognitive Deep Research
A cognitive deep research agent that emulates commercial analysts to synthesize expert-level reports via intent-driven search and iterative synthesis.
Why AI for Scientific Literature Mining Matters?
The explosion of scientific literature has exceeded human cognitive limits. We aim to build AI systems that can autonomously navigate, understand, and synthesize scientific knowledge, ultimately accelerating the pace of human discovery.
Autonomous Reasoning
Moving beyond keyword matching to deep semantic understanding and multi-hop reasoning across heterogeneous scientific documents.
Multimodal Synthesis
Integrating text, tables, figures, and equations to form a holistic understanding of scientific breakthroughs.
Hypothesis Generation
Discovering novel connections between disparate research areas to propose and validate new scientific hypotheses.