Tsinghua University, NVIDIA, CMU, UC Berkeley, Princeton and NUS have jointly launched StreamEval, the first streaming interactive video generation benchmark, covering five task types — digital human dialogue, virtual human performance, world generation, instruction editing and reference editing — with 300 test cases and 38 evaluated models. The code, dataset and scoring rules are fully open-sourced.
In response to OpenAI's release of 722 AI-generated mathematics manuscripts (three later withdrawn), mathematicians including Yau Shing-Tung, Xi Nanhua, He Xuhua, and Tian Ye said AI cannot replace human thinking and that cooperation norms are urgently needed. Yau called AI a tool for seeking truth that should be used freely, while Xi said OpenAI's lack of rules has caused significant difficulties for the academic community.
A Tsinghua-led team, together with multiple universities, released OSWorld-Science, a benchmark for computer-use agents in real scientific scenarios, featuring 146 expert-designed tasks across 7 scientific domains and 14 software configurations. Evaluated on 12 vision-language models, the average task score was 34.3%, with Claude Fable 5.1 leading at 73.7%.
On October 11, the Yau Mathematics and Artificial Intelligence Laboratory Alliance was established at Tsinghua University to promote deep integration of AI and mathematics, releasing over 10 key achievements and research directions in "AI+mathematics" involving scientific problems such as the degenerate Arnold conjecture, boundedness of minimal modifications, and 7-dimensional homotopy spheres, spanning geometry, quantum computing, and life sciences. AI mainly served as an assistant in exploring proof strategies, simplifying calculations, and providing inspiration.
On September 21, Tsinghua University, Infinigen, and Zhengxing Innovation jointly open-sourced RPent, an embodied agent infrastructure that connects general large-model task planning, VLA expert model fine manipulation, memory, tools, and robot interfaces to form a closed loop of perception, decision, execution, and feedback correction. It achieved a 92.6% task success rate on the LIBERO-PRO benchmark, improved end-to-end task completion speed by over 7 times, and released a series of real-robot demos.