Hugging Face Relaunches Papers with Code with Automated SOTA Leaderboards and Closed-Source Model Support
English summary
Niels from Hugging Face’s open-source team has relaunched paperswithcode.co as a platform to surface state-of-the-art results across AI domains by automatically parsing arXiv and Hugging Face papers. It generates interactive leaderboards with scatter plots and tables, illustrated by the BrowseComp benchmark. A key new feature is the inclusion of closed-source model evaluations (e.g., GPT-5.5, Mythos 5), treated as 'papers without code', with a toggle to show or hide them. The site also supports submissions from any source, not limited to preprint servers.
Chinese summary
Hugging Face 开源团队的 Niels 重新上线了 paperswithcode.co，通过自动解析 arXiv 和 Hugging Face 论文来展示各 AI 领域的当前最优结果，并生成可交互的排行榜、散点图与表格（以 BrowseComp 基准为例）。重要新增功能是纳入闭源模型评测（如 GPT-5.5、Mythos 5），将其视为“无代码论文”，用户可通过开关显示或隐藏。该平台还支持提交来自任意来源的结果，不限预印本。
Key points
Automatic parsing of arXiv/Hugging Face papers to create leaderboards across AI domains.
自动解析 arXiv 和 Hugging Face 论文，为各 AI 领域生成排行榜。
Closed-source model evals are supported, tagged as “papers without code,” with a toggle to show/hide them.
支持闭源模型评测，标记为“无代码论文”，并提供显示/隐藏开关。
Examples of closed-source entries include GPT-5.5 and Mythos 5.
闭源模型示例包括 GPT-5.5 和 Mythos 5。
Submission is open to any source (blog posts, etc.), not just arXiv.
接受任意来源的提交（博客等），不限于 arXiv。