Introduction to Paper Circle

Paper Circle is an open-source academic literature discovery and analysis framework based on multi-agent LLM, designed to address the problem of literature overload faced by contemporary researchers. Its core architecture adopts a dual-pipeline collaborative design of "discovery pipeline + analysis pipeline":

• Discovery Pipeline: Integrates multi-source academic resources, filters relevant literature through multi-criteria scoring and diversity ranking;
• Analysis Pipeline: Converts literature into structured knowledge graphs, supporting graph-aware question answering and coverage verification; The system is implemented based on a multi-agent orchestration framework and has been open-sourced (code repository: https://github.com/MAXNORM8650/papercircle). It is suitable for various scientific research scenarios such as literature review and new field exploration.

Background: Research Dilemmas and Opportunities for Solutions

The explosive growth of scientific literature has become a core challenge: PubMed adds over one million articles annually, and arXiv's daily submission volume is rising. Traditional keyword search and citation tracing are difficult to efficiently discover, evaluate, and synthesize research. The development of LLM and multi-agent systems provides new possibilities for solving this problem, and the Paper Circle project emerged as the times require, reducing the cognitive burden of researchers' literature research through automated agent collaboration.

Methodology: Dual-Pipeline Architecture and Multi-Agent Design

Dual-Pipeline Collaborative Architecture

1. Discovery Pipeline:
   • Multi-source retrieval: Integrates resources such as arXiv, PubMed, Google Scholar;
   • Multi-criteria scoring: Evaluation from relevance, influence, timeliness, and methodology dimensions;
   • Diversity ranking: Avoids the "echo chamber" effect and covers different research perspectives;
   • Structured output: JSON format includes metadata, scores, and other information.
2. Analysis Pipeline:
   • Knowledge graph schema: Nodes for concepts, methods, experiments, charts, claims, and their semantic relationships;
   • Graph-aware question answering: Cross-paper association queries;
   • Coverage verification: Identifies blind spots in literature reviews.

Multi-Agent Orchestration

• Role division: Retrieval, scoring, extraction, graph construction, and verification agents;
• Reproducible workflow: Standardized outputs (JSON, CSV, etc.) ensure process reproducibility.

Evidence: Performance Evaluation Results

Retrieval Performance

Test metrics include hit rate, Mean Reciprocal Rank (MRR), and Recall@K. Results show that as the agent model's capability increases, all metrics improve steadily, indicating the architecture has good scalability.

Review Generation Quality

Manual evaluation shows:

• Coverage and accuracy are close to the level of junior researchers;
• Coherence and critical analysis still need improvement.

Application Scenarios: User Value Across Multiple Scenarios

1. Literature Review Writing: Reduces retrieval and screening time, identifies research connections and differences through knowledge graphs;
2. New Field Exploration: Quickly builds a field map to understand core concepts, methods, and teams;
3. Research Trend Tracking: Runs the discovery pipeline regularly to automatically identify the latest progress and emerging directions;
4. Interdisciplinary Research: Multi-source retrieval and diversity ranking help discover interdisciplinary related work.

Limitations and Future Outlook

Current Limitations

• PDF parsing accuracy: The accuracy of information extraction from literature with complex layouts needs improvement;
• Knowledge graph completeness: Automatically built graphs may have omissions or errors and require manual review;
• Multilingual support: Mainly for English literature, with limited support for other languages.

Future Directions

• Enhance reasoning ability: Improve the planning and critical analysis capabilities of agents;
• Collaboration function: Support multi-user co-construction of knowledge graphs;
• Personalized recommendation: Provide customized literature recommendations based on user preferences;
• Real-time updates: Integrate with academic databases to realize real-time push of new literature.

Open-Source Ecosystem and Community Building

Paper Circle adopts a fully open-source strategy:

• Website: https://papercircle.vercel.app/
• Code repository: https://github.com/MAXNORM8650/papercircle

Open-source lowers the threshold for use and provides a foundation for community contributions and continuous improvement. Researchers can customize retrieval strategies, and developers can extend agent roles or knowledge graph schemas.