csnl-paper-scout: Multi-Agent Neuroscience Paper Recommendation System and Academic Workflow Automation

csnl-paper-scout, developed by the Cognitive and Systems Neuroscience Laboratory (CSNL) at Seoul National University, is an AI-driven paper recommendation pipeline. It automatically discovers relevant papers via semantic embedding RAG, uses a five-dimensional value scoring system and a six-agent team to collaboratively generate recommendation content, and finally publishes to Slack, enabling end-to-end automation of academic information and solving the problem of manual screening of massive literature.

With the explosive growth in the number of published papers, manual screening of relevant literature in specific research directions (such as cognitive and systems neuroscience) has become increasingly difficult. CSNL developed csnl-paper-scout to precisely discover papers highly relevant to the lab's research through AI technology, improving literature tracking efficiency.

The system uses a five-stage pipeline: Scan→Score→Team→Review→Post:

1. Scan Stage: 90-day rolling window, with dual filtering via Zotero similarity ≥0.4 and embedding vector similarity ≥0.45;
2. Score Stage: Evaluation across five dimensions (direct relevance, theoretical conflict/supplement, methodological reference, competitive comparison, problem re-framing), taking the maximum value as the comprehensive score;
3. Team Drafting Stage: Collaboration among six agents (drafter, hook evaluator, visual agent, accuracy evaluator, member spokesperson, final editor), supporting 3 rounds of feedback loops;
4. Review Stage: Optional group peer review for borderline papers with scores of 7-8;
5. Post Stage: Publish recommendation content (Korean + English citations) to the Slack #study-paper-reading channel, which requires user confirmation before posting.

The system relies on the following configuration and data modules:

• Central configuration (context-bundle.json): Contains lab member information, research projects, etc.;
• Reading database (reading-db/): Records paper reading history, deduplicates to avoid repeated recommendations;
• Member profiles: Automatically generates interest profiles (focused topics, tracked authors, reading history);
• Auxiliary configurations such as collaboration networks, journal rankings, and tracked author lists.

The system's effectiveness is evaluated through seven dimensions:

Dimension	Name	Assessment Type
E1	Structural Compliance	Automated
E2	Semantic Fidelity	LLM Assessment
E3	Member Targeting	LLM Assessment
E4	Hook Effectiveness	Automated
E5	Agent Convergence	Automated
E6	Pipeline Coherence	LLM Assessment
E7	Safety	Automated (hard failure gatekeeping)
The rating levels are A (≥0.85), B (≥0.70), C (≥0.55), D (≥0.40), F (<0.40 or E7 failure).

The core innovations of the system include:

1. Quantitative Relevance Assessment: The five-dimensional scoring system converts vague relevance into actionable metrics;
2. Multi-Agent Collaboration: Six agents divide labor and iteratively optimize recommendation content;
3. Closed-Loop Learning: Synchronously update the reading database and member profiles via Slack messages to continuously optimize recommendations;
4. Human-Machine Collaboration Boundary: User confirmation is required before publishing, respecting human decision-making rights.

The system design can be extended to:

• Other academic fields: Adapt by adjusting research topics and journal lists;
• Enterprise technology tracking: Replace neuroscience journals with technical blogs/arXiv computer science categories;
• Personal literature management: Simplify the multi-agent process as a personal assistant;
• Open-source community maintenance: Track progress in specific technical fields and push updates.

csnl-paper-scout is a customized solution deeply integrated with CSNL's research direction, balancing automation efficiency and recommendation quality. Its modular design provides a complete technical reference for other labs, allowing for tailoring and expansion according to needs, making it a typical practice in academic research automation.