The authors constructed ANCHOR from videos spanning social interactions, egocentric perception and instructional tasks.
Benchmark exposes when AI reasoning ignores audio-visual evidence
ANCHOR measures whether multimodal LLMs cite the right visual, spatial and audio cues, revealing large grounding gaps even among models with similar answer accuracy.
Big Tech
Joel Julin · Souraja Kundu · Liza Dahiya · George Z. Wei · Liuyue Xie · Jingru Yi · +3 more
Carnegie Mellon University · Amazon
Research Digest··3 min read
The authors introduce ANCHOR, a benchmark of 2,319 videos and 8,272 question-answer pairs with structured reasoning traces, temporal evidence, bounding boxes and audio cues.
§