jusjinuk/daily-papers

Daily Papers Review

When to use

• User asks to check today's papers, do a daily paper review, or says "daily papers"
• Triggered by the scheduler (e.g., cron every weekday morning)
• User asks "any interesting papers today?"

Research Topics

Default topics (pick the paper most relevant to any of these):

1. LLM efficiency (quantization, pruning, distillation, inference optimization)
2. Reinforcement learning (RLHF, RLAIF, reward modeling, policy optimization)
3. Continual learning for LLM agents (lifelong learning, memory, catastrophic forgetting)
4. Diffusion LLM (discrete diffusion, text generation via diffusion, hybrid architectures)
5. Time series forecasting (foundation models for time series, multivariate forecasting)

If the user specifies different topics (e.g., "focus on RL today"), use those instead.

Papers Log

A log of previously reviewed papers is kept at reports-repo/papers-log.json to avoid re-reviewing.

Format:

[
  {"arxiv_id": "2506.12345", "title": "Paper Title", "reviewed_on": "2026-03-04", "report": "2026-03-04-daily-papers.pdf"}
]

If the file doesn't exist, create it as an empty array [].

Workflow

1. Fetch today's papers

Use mcp__hf_papers__get_today_papers to get the day's trending papers.

2. Filter and select

• Read reports-repo/papers-log.json — collect all previously reviewed arxiv_id values
• Skip any paper whose arxiv ID is already in the log
• From the remaining papers, pick the 1 paper most relevant to the research topics
• If the user overrode topics, use those instead of defaults
• If no papers match any topic, post to Slack: "No new papers matching your research topics today." and stop

3. Study the paper

• Use mcp__arxiv__get_paper to fetch the full paper metadata
• Download and read the PDF if possible using WebFetch
• Deeply analyze the paper:
  • What is the core problem and why does it matter?
  • What is the key technical contribution?
  • What are the main results (quantitative where possible)?
  • What are the limitations and assumptions?

4. Brainstorm follow-up research

Generate 2-3 concrete follow-up research directions:

• Each must be specific enough to start working on (not vague like "improve performance")
• For each direction include:
  • One-line description
  • What gap it addresses
  • Approach sketch (2-3 sentences)
  • Feasibility: easy (weeks) / moderate (months) / hard (6+ months)
• Prioritize by intersection of novelty and feasibility

5. Write LaTeX report

Write a report to reports-repo/reports/YYYY-MM-DD-daily-papers.tex with:

• \section{Paper Summary} — title, authors, core contribution, method, results, limitations. Include links:
  • arxiv: https://arxiv.org/abs/{id}
  • HF daily papers page: https://huggingface.co/papers/{id}
  • GitHub repo (if the paper has an open-source implementation — check the paper metadata or HF page for a repo link; omit if none exists)
• \section{Follow-up Research Directions} — the 2-3 directions with full detail
• \section{Conclusion} — key takeaway and recommended next step
• Proper \bibliography with BibTeX entries for the paper and any referenced work

Compile:

cd reports-repo/reports
pdflatex -interaction=nonstopmode YYYY-MM-DD-daily-papers.tex
bibtex YYYY-MM-DD-daily-papers
pdflatex -interaction=nonstopmode YYYY-MM-DD-daily-papers.tex
pdflatex -interaction=nonstopmode YYYY-MM-DD-daily-papers.tex

6. Publish

• From reports-repo/, commit the .tex, .bib, and .pdf files
• Push to GitHub
• Update reports-repo/papers-log.json — append the new entry, commit and push

7. Post to Slack

Post a concise summary (not the full report):

• Paper title with links: arxiv, HF daily papers page, GitHub repo (if exists)
• 2-3 sentence summary of what the paper does and why it matters
• Top follow-up research idea (1 sentence)
• Link to the PDF report: get the repo URL via git -C reports-repo remote get-url origin, then construct <repo-url>/blob/main/reports/YYYY-MM-DD-daily-papers.pdf (strip any .git suffix and auth prefix)