LLM4ArxivPaper

中文文档

LLM4ArxivPaper crawls arXiv for papers in your research areas, runs an LLM analysis tailored to your interests, and serves the results through a web app and a weekly email. It has two parts: an analysis pipeline that runs on GitHub Actions, and a Next.js app deployed on Vercel.

Features

• Works with any OpenAI-compatible endpoint (DeepSeek / OpenAI / Bailian / local vLLM); paper bodies come from ar5iv, so token cost stays low
• Relevance note: one or two sentences on how each paper connects to your research
• Five-aspect analysis (problem / solution / methodology / experiments / conclusion) plus interest-driven Q&A with quoted evidence
• Key figures extracted automatically (architecture and flow diagrams)
• On-demand analysis from a URL, plus chat with the paper's full context, history synced across devices
• Star papers you want to keep; hide the ones you're not interested in
• Backfill any historical date range
• A heartbeat commit dodges GitHub's 60-day auto-disable of scheduled workflows

Screenshots

The home page lists every analysed paper, with tabs for all papers, the current week's run, and the ones you starred:

Each paper page opens with the relevance note (why this paper matters to your work) and the at-a-glance summary:

Below that is the five-aspect analysis. The method/architecture figure is pulled straight from the paper and embedded inside the methodology aspect:

Install

# 1. Create your own private repo with this repo's "Use this template", then clone it
# 2. Validate your LLM config locally (~1-2 min)
pip install -r requirements.txt
API_KEY="sk-..." BASE_URL="https://api.deepseek.com/v1" \
  python src/main.py run --paper-limit 1

--paper-limit 1 analyses one paper per topic. The terminal shows the crawl, scoring, and five-aspect analysis; a JSON file lands under data/analyses/. If this works, your LLM key, endpoint, and topic config are correct.

Then configure on GitHub:

1. Settings → Secrets and variables → Actions: add API_KEY and BASE_URL.
2. Settings → Actions → General: allow workflows to run.
3. Edit config/pipeline.yaml with your research areas:

openai:
  api_key: "${API_KEY}"
  base_url: "${BASE_URL}"
  relevance_model: "deepseek-v4-flash"
  summarization_model: "deepseek-v4-flash"
  language: "zh-CN"          # analysis output language: zh-CN or en

topics:
  - name: "software_testing"
    label: "Software Testing"
    query:
      categories: ["cs.SE", "cs.AI"]
      include: ["test generation", "test update"]
      exclude: ["quantum"]
    interest_prompt: |
      I research LLM-assisted software testing, focused on test-case
      generation and fault localisation. I want to understand how LLMs
      and static analysis can be combined effectively.

4. Actions → Weekly LLM4ArxivPaper Pipeline → Run workflow for the first full run.

The pipeline writes data/analyses/*.json, committed to your repo. Those JSON files are the data source for the web app and the email. language controls the language of everything the LLM generates, applied at generation time.

Deploy the web app

The web app provides browsing, on-demand URL analysis, chat, starring, and hiding papers you don't care about. It runs on Vercel.

1. Sign up for Vercel and import your repo as a project.
2. Project Settings → General → Root Directory: set to web.
3. Project Storage → Browse Marketplace → Upstash Redis → Install. Redis holds stars and chat history; its connection env vars are injected automatically.
4. Create a GitHub fine-grained PAT with Actions: Read and write, Contents: Read and write, Metadata: Read, scoped to your instance repo only.
5. Project Settings → Environment Variables:

Variable	Value	Required
LLM_API_KEY	Same as API_KEY	yes
ADMIN_TOKEN	A long random string you choose, used to log in	yes
GH_DISPATCH_TOKEN	The PAT from the previous step	yes
LLM_BASE_URL	LLM endpoint	no, defaults to DeepSeek
LLM_MODEL	Model id	no, defaults to deepseek-v4-flash
ARCHIVE_CHATS	true if your repo is private	no, defaults to false

The Redis connection vars are injected by step 3; the repo owner / name are derived from Vercel's git integration.

Vercel auto-deploys on push. Open the deployment URL, go to /login and paste your ADMIN_TOKEN, then /submit and paste an arXiv link; the analysis page loads in ~2-3 minutes.

Email digest

Get the weekly picks in your inbox.

1. Enable two-factor auth on your Google account and generate a Gmail app password.
2. Repo Settings → Secrets and variables → Actions: add MAIL_USERNAME (Gmail address) and MAIL_PASSWORD (app password).
3. Set the recipient in config/pipeline.yaml:

email:
  enabled: true
  recipients: ["you@example.com"]

MAIL_* is also used to email you an alert if the pipeline fails. For another provider, change email.smtp_host / smtp_port.

How it works

Three GitHub Actions workflows:

Workflow	Trigger	What it does
weekly-pipeline.yml	Mon 02:00 UTC + manual	Crawl, score, analyse, commit data/, send email, and commit a heartbeat file to stay alive
backfill.yml	Manual	Crawl a date range in chunks for historical papers; dry_run=true previews count and cost first
analyse-one.yml	Triggered by the web app's /submit, or manual	Fetch and analyse a single paper

Every fetched paper is scored 0-100 for relevance against your interest_prompt. Only papers at or above relevance.pass_threshold (default 60) get a full deep analysis; the rest are dropped, so the LLM budget goes to papers that actually match your interests. Papers you submit yourself via /submit are always analysed and still show their honest score.

Storage is the repo itself: analyses are JSON files under data/, no external database. Live state like stars and chats lives in Upstash Redis, snapshotted back into the repo daily by a cron for backup.

Customising

Change the weekly schedule

Edit the cron expression at the top of .github/workflows/weekly-pipeline.yml (fields are minute hour day month weekday, time is UTC):

on:
  schedule:
    - cron: '0 2 * * 1'    # Mondays at 02:00 UTC

Goal	cron
Mondays 8am Beijing time (UTC+8)	0 0 * * 1
Every morning	0 0 * * *
Mondays and Thursdays	0 2 * * 1,4

You can also hit Run workflow in the Actions tab any time.

Backfill past papers

Go to Actions → Backfill arXiv Papers → Run workflow:

• start_date / end_date: the date range, YYYY-MM-DD.
• chunk_days: days per slice (default 7).
• paper_limit: max papers analysed per topic per slice, to cap cost.
• dry_run: set true the first time. It only counts candidate papers, calls no LLM, costs nothing. Once you've seen the count, set it false for the real run.

A real run commits the backfilled analyses to data/; papers already analysed are skipped.

Tune fetching and filtering

In config/pipeline.yaml:

Field	What it does
fetch.days_back	How many days back the weekly run looks (default 7)
fetch.max_papers_per_topic	Max candidate papers fetched per topic
relevance.pass_threshold	Relevance cutoff; papers below it get no deep analysis (default 60)
summarization.max_content_chars	Max paper-body characters sent to the LLM
topics	Add or remove research areas, each with its own query and interest_prompt
email.recipients	Email recipient list

Privacy

Your instance repo holds your analyses, stars, and chat history. Keep it Private. Live state is in your own Upstash account. None of it passes through a third-party service.

The Vercel deployment is a public URL; anyone with the link can browse it. Write actions (submitting papers, starring) require the ADMIN_TOKEN login, so unauthenticated visitors are read-only. To restrict browsing too, use Vercel's Deployment Protection.

For a public instance, ARCHIVE_STARS / ARCHIVE_CHATS control what gets snapshotted into git: stars are archived by default, chats stay in Redis only.

Cost

With DeepSeek V4 Flash, each paper analysis runs about $0.01–0.05; a weekly run of 20-40 papers stays well under $1. The free tiers of GitHub Actions, Vercel, and Upstash are ample for a single-user instance. Heavy compute is on GitHub Actions (6-hour job limit); Vercel only does the lightweight UI and API.

FAQ

My scheduled workflow stopped after a couple of months. GitHub auto-disables scheduled workflows after 60 days of repo inactivity. The weekly run commits a heartbeat file to dodge that. If yours is already disabled, re-enable it once from the Actions tab; the heartbeat keeps it alive afterwards.

/submit returns 403 Resource not accessible by personal access token. The PAT is missing Actions: Read and write. Fine-grained PATs can't be edited after creation; regenerate one with the correct permissions.

Can I use it without Vercel? Yes. Install plus email digest is a complete no-Vercel path: the pipeline analyses papers weekly and the picks land in your inbox. You give up chat, on-demand URL analysis, and starring.

A paper has no key figures. Figures come from ar5iv's HTML rendering. Papers ar5iv hasn't rendered (too new, or no LaTeX source) degrade to no figures; the rest of the analysis is unaffected.

Project layout

src/                       Python pipeline (runs on GitHub Actions)
  fetchers/                arXiv API + ar5iv HTML / figure / PDF extraction
  filters/                 LLM relevance scoring
  summaries/               the core analysis pipeline
  storage/                 writes data/*.json
  publisher/               weekly email digest
  workflow/                CLI and orchestration
web/                       Next.js app (runs on Vercel)
config/pipeline.yaml       research areas + interests + model config
data/                      analyses, stars, chats (JSON files)
.github/workflows/         the four workflows
docs/                      README_zh.md and any other docs
pics/                      screenshots and the workflow diagram

Contributing

Issues and PRs welcome. The contract between the Python and TypeScript halves is the JSON shape written by _summary_to_payload in src/workflow/pipeline.py and read by web/lib/data-reader.ts; change fields on both sides together.

License

MIT, see LICENSE.