Researchers have introduced MotionHalluc, a new benchmark designed to identify and diagnose kinematic hallucinations in large multimodal models that perform motion reasoning. The benchmark consists of 1540 fine-grained questions across 553 video pairs, evaluating directional, attributional, and temporal hallucinations. Experiments show that current state-of-the-art models are highly susceptible to these issues. To combat this, the team also proposed Perceive-Parse-Verify (PPV), a training-free method that injects explicit quantitative measurements, leading to an average performance gain of 10.6% across evaluated models. AI
IMPACT Highlights a specific failure mode in AI motion reasoning, potentially guiding future model development for improved accuracy in video analysis.
RANK_REASON The cluster describes a new academic paper introducing a benchmark and a method for evaluating AI models. [lever_c_demoted from research: ic=1 ai=1.0]
- alphaXiv
- arXiv
- CatalyzeX
- DagsHub
- Gotit.pub
- Hugging Face
- MotionHalluc
- Perceive-Parse-Verify
- ScienceCast
- Weile Guo
AI-generated summary · Google Gemini · from 1 sources. How we write summaries →
