Hugging Face AI Updates: May 25, 2026

1. Hugging Face relaunches PapersWithCode with multi-metric leaderboards, non-arXiv submissions, and roughly 3,000 evals

Hugging Face. Niels Rogge from the Hugging Face open-source team announced a revived paperswithcode.co on May 24, expanding state-of-the-art tracking beyond the original site’s scope across domains including agents, computer vision, time-series forecasting, automatic speech recognition, and object detection. Leaderboards now carry multiple metrics per benchmark so the Open ASR board surfaces both Word Error Rate and Inverse Real-Time Factor, and the COCO object detection board reports mean Average Precision alongside frames-per-second, making throughput-versus-quality tradeoffs visible at a glance. The submission portal at paperswithcode.co/submit accepts papers from sources outside arXiv such as GitHub repos, blog posts, and BiorXiv (DeepSeek-v4 is cited as an example), with AI-assisted enrichment auto-tagging tasks, methods, repos, and evaluations on intake. Other additions include paper lineage banners that display follow-ups and predecessors on entries like Mamba-3, DINOv2, and GLM-4.5, an expanded methods library covering Gated DeltaNet, Kimi Delta Attention, and Mamba-2 with full citing-paper lists, copy-image buttons on scatter plots and leaderboard tables for direct social sharing, and an evaluation database of approximately 3,000 evals that covers every model supported in the Transformers library and renders at the bottom of paper pages. Source