ArXiv Paper Summarizer

A GitHub Actions workflow that fetches recent ArXiv papers, filters them using LLMs, summarizes them, and publishes the results as GitHub Issues.

Features

• Smart Filtering: Uses gpt-5-mini to score paper relevance (0-10) based on your keywords.
• Concise Summaries: Generates high-quality summaries using gpt-5.
• Flexible LLM Support: Uses Microsoft Foundry/Azure OpenAI by default and supports other OpenAI-compatible APIs.
• Incremental Fetching: Only fetches papers published since the last run to avoid duplicates.
• Daily Schedule: Runs automatically every day at 07:00 UTC.
• Reading List: Mark papers to read later with a checkbox, automatically tracked in a dedicated issue.

Configuration

Edit config.yaml to customize your preferences:

arxiv:
  categories:
    - "cs.AI"
    - "cs.LG"
  keywords:
    - "LLM"
    - "Agent"
  max_results: 20

github:
  usernames:
    - "your-username" # Users to tag in the issue
  issue_label: "arxiv-summary"
  max_papers_per_issue: 10 # Split into multiple issues when more papers are found

llm_service:
  # Set this here or with LLM_BASE_URL:
  base_url: "https://<resource>.openai.azure.com/openai/v1/"

models:
  # Azure deployment names
  filter: "gpt-5-mini"
  summarize: "gpt-5"

Microsoft Foundry / Azure OpenAI

This repository consumes pre-existing Azure resources; it does not provision them.

1. Create a Microsoft Foundry or Azure OpenAI resource.
2. Deploy gpt-5-mini for filtering and gpt-5 for summarization. If you choose different deployment names, update the models values in config.yaml.
3. Copy the resource's OpenAI-compatible v1 endpoint and API key.

Supported endpoint formats include:

llm_service:
  base_url: "https://<resource>.openai.azure.com/openai/v1/"

# A Microsoft Foundry resource endpoint can also be used:
llm_service:
  base_url: "https://<resource>.services.ai.azure.com/openai/v1/"

Model availability and quota vary by Azure region.

Using Other LLM Providers

The tool retains support for OpenAI-compatible APIs. Configure llm_service in config.yaml or set LLM_BASE_URL:

# OpenAI
llm_service:
  base_url: "https://api.openai.com/v1"

# Local
llm_service:
  base_url: "http://localhost:11434/v1"

Set the provider configuration through environment variables:

• LLM_BASE_URL - Overrides the base URL in config.yaml.
• LLM_API_KEY - The provider API key.

GitHub Actions

ArXiv Summarizer

The workflow is defined in .github/workflows/summarize.yml. It is configured to run:

• Daily at 07:00 UTC.
• Manually via the "Run workflow" button in the Actions tab.

Configuring Microsoft Foundry in GitHub Actions

Add these repository secrets before running the workflow:

1. LLM_BASE_URL - The Azure OpenAI-compatible v1 endpoint.
2. LLM_API_KEY - The Azure resource API key.

The workflow's GITHUB_TOKEN is used only to read and create GitHub Issues.

Reading List

Each paper summary includes a "📚 Read Later" checkbox. When you check this box:

1. A GitHub workflow detects the change (.github/workflows/reading-list.yml)
2. The paper title and ArXiv link are automatically added to a 📚 ArXiv Reading List issue
3. The reading list issue is created automatically if it doesn't exist (labeled reading-list)

This lets you quickly bookmark interesting papers while reviewing the daily digest, with all your selections tracked in one place.

Local Development & Testing

You can run the tool locally without GitHub Actions.

Prerequisites

1. Install uv (or use pip).
2. A GitHub Personal Access Token (PAT) with issue read/write access.
3. A Microsoft Foundry/Azure OpenAI v1 endpoint, API key, and deployed models.

Setup

# Install dependencies
uv sync

Running Locally

1. Create a .env file in the root directory:

   GITHUB_TOKEN=your_fine_grained_token
   GITHUB_REPOSITORY=owner/repo
   LLM_BASE_URL=https://your-resource.openai.azure.com/openai/v1/
   LLM_API_KEY=your_azure_api_key

2. Run the summarizer:

   uv run src/main.py

Token Permissions (Fine-grained):

• Issues: Read and Write (to create summaries and check last run).

Note

When running locally, the tool will try to create a real issue in the specified repository. If you just want to test the fetching/summarizing logic without creating an issue, you can modify src/main.py or src/issue_creator.py temporarily.