Paper Radar – daily ArXiv triage, now runs on Claude Haiku too
Paper Radar, an open-source tool, uses AI to evaluate every new paper on arXiv each morning, filtering thousands of submissions down to relevant research for just $0.06 per day.
arXiv now receives over 30,000 papers monthly, with approximately 1,500 published each weekday. Traditional methods like keyword alerts and embedding recommenders frequently miss relevant papers that use different terminology or don't resemble past readings. Paper Radar solves this by evaluating every new paper using AI judgment rather than generating summaries, making it comprehensive and affordable.
The tool uses TypeSafe's System One model (Jev) to ask a simple yes/no question per interest and returns a probability score. This approach costs significantly less than generating summaries—about $0.06 daily to process all new arXiv papers. The system has been tested against real reviewers, showing it can reproduce 96.9% of included studies while reducing reading load by 78%.
How It Works
Users fork the repository, describe their interests in plain English, and add an API key. The system then runs daily, evaluating papers and providing results via GitHub Pages, RSS feeds, and optional notifications. Each decision and probability is logged transparently, allowing users to verify the process.
Judgement instead of prose is what makes the arithmetic work. There is nothing to parse and no invented labels, and one judgement costs a fraction of a generated sentence — cheap enough to read all 1,500 papers rather than a pre-filtered handful.
Alternative Models
While the default uses TypeSafe's Jev model, users can alternatively configure Paper Radar to use Anthropic's Claude through their API key. This option provides the same pipeline and calibration step without requiring a waitlist for TypeSafe's service. The tool also supports OpenRouter as another backend option.
The project is designed to be self-hosted on GitHub Actions, requiring no server setup. Users can customize their radar configuration to focus on specific research areas, making the tool more relevant to their needs. The system also provides optional one-sentence summaries in the user's language for the top 10 papers only.
Related
