Hugging Face revives PapersWithCode with AI-powered leaderboards
Niels rebuilds the dead platform using AI agents to parse papers at scale.
Niels Rogge from Hugging Face's open-source team has launched paperswithcode.co, a revival of the beloved PapersWithCode platform that Meta abandoned after its acquisition. The new version leverages AI agents to parse papers at scale, automatically generating leaderboards for state-of-the-art models. Already parsed are high-impact works like Qwen 3.5 and 3.6, RF-DETR for object detection, DINOv3, MTEB embedding model results, and the Open ASR Leaderboard for speech recognition. The site organizes papers by domain and shows trending papers based on GitHub star velocity, plus citation counts for most-cited works.
Features include automated linked GitHub repos, project pages, and artifact storage (with multiple repos per paper). It supports external papers beyond arXiv—for example, DeepSeek v4 harness reports for coding agent benchmarks like Terminal Bench 2.0. Users can sign in with Hugging Face accounts, while Storage Buckets store thumbnails, PDFs, and data backups. Niels is actively verifying results manually and invites feedback and feature requests from the community.
- Uses AI agents to parse papers and auto-generate SOTA leaderboards across domains
- Currently covers papers like Qwen 3.5/3.6, RF-DETR, DINOv3, MTEB embeddings, and Open ASR
- Supports citation counts, trending by GitHub stars, external papers, and multiple repos per paper
Why It Matters
Restoring a centralized hub for state-of-the-art benchmarks keeps AI research discoverable and reproducible.