Benchmark RadarMotionHalluc
MotionHalluc is a benchmark evaluating kinematic hallucinations in cross-video motion comparison. It includes 1540 questions over 553 video pairs, assessing directional, attributional, and temporal hallucinations in generated instructions.
- Released
- 2026-06-22
- Readiness
- Paper only
- Primary field
- General AI
Why it matters
Large multimodal models often produce motion hallucinations in paired-video comparison tasks. A systematic benchmark like MotionHalluc could help measure and reduce these errors, potentially improving the reliability of automated motion feedback systems.
Motivation
Motion instruction generation in cross-video comparison aims to produce corrective feedback that describes the differences between a query and a reference motion.
Primary resources
Benchmark Radar records only publicly supported details and links back to primary sources for verification.