The authors built Omni-InteractScope from 1,479 human-annotated, cross-reviewed interactions containing 1,790 temporal targets.
Streaming AI assistants struggle to act on time and sustain guidance
Across 12 models, Omni-InteractScope finds low rates of timely tool use, unsolicited reminders and complete multistep guidance in continuous audiovisual streams.
Chinese Tech
Yinsong Yan · Ling Wang · Yuxuan Wang · Yifan Yang · Ruiyang Xu · Junming Lin · +8 more
The Hong Kong Polytechnic University · Alibaba Token Hub, Alibaba Group · Shanghai Jiaotong University · Tsinghua University · Fudan University
Research Digest··2 min read
Yan et al. introduce a benchmark testing whether AI assistants watching and listening to continuous streams intervene when needed and keep responding as events unfold.
The Daily Edition
The front page in your inbox each weekday morning: the lead, the picks, and why each one earned its place.