Researchers have introduced HanDyVQA, a new video question-answering benchmark designed to evaluate fine-grained understanding of hand-object interaction dynamics. The benchmark includes over 11,000 QA pairs across six question types, focusing on manipulation styles, motion, and part-level state changes. Even advanced models like Gemini 2.5 Pro struggled, achieving only 73% average accuracy compared to human performance of 97%, highlighting ongoing challenges in spatial relationship and geometric understanding. AI
IMPACT Highlights limitations in current video foundation models for understanding complex human-object interactions, guiding future research.
RANK_REASON The cluster describes a new academic benchmark for evaluating AI models on a specific task, published on arXiv. [lever_c_demoted from research: ic=1 ai=1.0]
AI-generated summary · Google Gemini · from 1 sources. How we write summaries →