The authors constructed AV-FABLE, a benchmark of 16,650 pristine–manipulated twin video pairs totaling 998 hours, with 27 editing configurations covering sparse manipulations of speech, sounds, actions, and objects in open-scene long-form videos.
Benchmark reveals AV-LLMs encode but fail to verbalize manipulation evidence
AV-FABLE provides 16,650 pristine–manipulated video pairs with separate audio and visual timelines; supervised readouts of frozen features outperform zero-shot prompting.
Top University
Yuesheng Huang · Hoi Fan Au · Jiaqi Hou · Ruicheng Zhang · Jiarui He · Haocheng Yang · +5 more
Tsinghua University · Peking University
Research Digest··3 min read
The authors introduce AV-FABLE, a dual-timeline benchmark for detecting sparse audio-visual manipulations in long-form videos.
The Daily Edition
The front page in your inbox each weekday morning: the lead, the picks, and why each one earned its place.