Hugging Face revives Papers With Code with AI-powered benchmark parsing
Hugging Face has relaunched Papers With Code at paperswithcode.co, using AI agents to automatically parse research papers and generate leaderboards after Meta left the original site dormant.
Papers With Code is live again. Niels Rogge from Hugging Face's open-source team has rebuilt the research aggregator at paperswithcode.co, filling the gap left when Meta acquired and abandoned the original platform. The new version uses AI agents to parse papers at scale and automatically generate benchmark leaderboards; Rogge manually verifies results for now.
The relaunch focuses on high-impact papers with well-established state-of-the-art claims—Qwen 3.5 and 3.6, RF-DETR for object detection, DINOv3, top embedding models from the MTEB leaderboard, and speech recognition models from the Open ASR Leaderboard. Each paper page surfaces evaluation results, linked GitHub repositories, project pages, and artifacts. The site defaults to trending papers ranked by GitHub star velocity and supports filtering by citation count across domains. Domain-specific leaderboards include MMTEB and COCO val 2017, plus Harness reports for coding agent benchmarks like Terminal Bench. Methods pages catalog techniques like RLVR, restoring a feature from the original site.
The platform indexes papers beyond arXiv—DeepSeek v4 and other external releases are included. Authentication uses Hugging Face accounts, and the backend stores thumbnails, PDFs, and backups in HF Storage Buckets. The immediate challenge is scaling the AI parsing pipeline beyond manually curated high-impact papers to cover the long tail of arXiv uploads without flooding leaderboards with unverified claims. Rogge is soliciting feedback and feature requests from the community.