AstroPaper: An AI-Powered arXiv Paper Tracking and Reading Workflow

AstroPaper is an open-source LLM agent workflow system designed to help astronomy researchers automatically track latest papers in arXiv's astro-ph.GA field. It filters papers based on personal research interests, generates summaries, and produces structured reading reports. The project offers dual value: a continuously updated reading log for specific astrophysics topics (high red shift galaxies, IGM/CGM, quasar environments, cosmic web, reionization) and reusable LLM agent workflow templates adaptable to other research fields. Its core idea is to use LLM agents to automate paper screening and note-taking while preserving researchers' decision-making authority.

Active researchers face challenges in keeping up with field progress. For example, arXiv's astro-ph.GA category releases dozens of new papers daily. Traditional manual browsing is time-consuming and prone to missing key works; social media or email list recommendations lack systematics and personalization. AstroPaper addresses this by leveraging LLM agents for automated processing.

AstroPaper's workflow involves three steps:

1. Define Research Interest: Copy doc/template_research_profile.md to doc/research_profile.md (gitignored for privacy) and edit to specify focus topics, keywords, paper types (theory/observation/simulation), research objects (redshift range, galaxy types), and reading priorities.
2. Delegate Agent Execution: Use LLM programming agents (Claude Code, Codex CLI) with simple commands (e.g., "check papers from last week"). The agent reads workflow docs, determines time windows, scans arXiv, matches papers to user profile, generates structured reports, and saves/renders outputs.
3. Review & Use: Reports are stored in year directories (e.g., arXiv_20260417w.md for weekly reports) and include weekly overview, selected papers (detailed summaries/comments), quick browse list, and research implications.

AstroPaper enforces strict report standards:

• Three-layer Summary: For each paper: 1-sentence core finding, paragraph overview (background/method/results), detailed notes (method details, data features, comparisons, limitations).
• Source Priority: Label sources as "source" (original paper, highest), "skim" (secondary), "mention" (background).
• Writing Style: Objective, distinguish author claims from agent inferences, note uncertainties (e.g., "paper claims..." vs "data shows..."), formal scientific tone.

AstroPaper's key contribution is transferability:

• Fork & Adapt: Clone the repo to other fields (cond-mat, cs.CL, hep-th), modify research_profile.md to define new scopes—workflow remains unchanged.
• Customize Workflow: The doc/agent_workflow.md manual includes directory structure, report frequency codes (w/weekly, d/daily, s/special), summary formats, source priority rules, and style guides—researchers can adjust any part for personalized workflows.

Technical Implementation:

• Uses LLM programming agents (Claude Code, Codex CLI, GPT with browsing) that learn from workflow docs via context learning (no fine-tuning needed).
• Outputs: Markdown (version control-friendly) + optional PDF (via scripts/render_report_pdf.sh using pandoc/xelatex).
• Version control: Dual license (MIT for workflow/scripts, CC BY 4.0 for reports); year-based directories; gitignored research profile for privacy.

Academic Integrity:

• Disclaimer: Notes are not peer-reviewed—verify original papers before citing.
• Human-in-loop: Agent handles info collection/initial sorting; researchers retain final judgment.
• Traceability: Reports label the LLM model used (e.g., "Working model: Claude Code") for reproducibility.

Applications:

• Literature review preparation (systematic tracking avoids missing key papers).
• Research inspiration (cross-subfield reading sparks method transfer ideas).
• Teaching (structured notes aid graduate students in learning the frontiers of the field).
• Cross-disciplinary exploration (easily create logs for multiple fields).

Limitations:

• Model hallucination risk: LLM may misinterpret content—verify critical info manually.
• Selective bias: Research profile defines focus, potentially creating information cocoons—update profile regularly.
• Depth vs breadth: Automation covers breadth but deep understanding requires manual effort.
• Technical threshold: Needs API access, command-line skills to use LLM programming agents.

AstroPaper is a well-designed academic tool that integrates LLM agents into daily workflows. It automates info processing to free researchers for high-value creative work, without replacing their judgment. For efficiency-seeking researchers, it's a ready-to-use starting point; for AI-in-research observers, it's a valuable case study.

Key implications:

• From passive tools to active partners: AI agents take on cognitive-heavy tasks (information filtering, initial analysis).
• Balance personalization & standardization: "Research profile + standard workflow" ensures consistency and customization.
• Open-source collaboration: Workflows become shared assets—community can optimize prompts, templates, etc.