MotionHalluc: Diagnosing Kinematic Hallucinations in Fine-Grained Motion Reasoning

Researchers introduce MotionHalluc, a benchmark dataset for evaluating how AI models hallucinate when analyzing motion differences between paired videos. The study reveals that large multimodal models struggle with directional, attributional, and temporal hallucinations in motion reasoning, but shows that injecting explicit kinematic measurements can improve accuracy by 10.6%.

MotionHalluc addresses a critical limitation in multimodal AI systems: their tendency to generate plausible-sounding but factually incorrect descriptions when comparing videos. This matters because motion understanding underpins applications from sports analytics to physical rehabilitation feedback systems, where accuracy directly impacts user outcomes. The benchmark's three-dimensional hallucination taxonomy—directional, attributional, and temporal—provides researchers with a structured framework for understanding failure modes rather than treating hallucinations as monolithic problems.

The research builds on growing recognition that large vision-language models excel at pattern recognition but struggle with precise quantitative reasoning. Previous work identified hallucinations in image captioning and visual question-answering; this study extends that concern specifically to kinematic understanding, a domain requiring both visual perception and physics-aware inference. The introduction of paired-video comparison as a testing ground reflects real-world applications where systems must explain differences rather than describe individual observations.

The Perceive-Parse-Verify framework's modest but consistent 10.6% performance gain suggests that explicit measurement injection—converting qualitative instructions into quantifiable metrics—bridges a fundamental gap in how AI systems approach motion analysis. This finding has implications for any AI system requiring precise physical reasoning, from autonomous systems to medical motion analysis. The training-free nature of PPV makes it immediately applicable to existing models.

Future developments will likely focus on whether measurement-grounded training approaches yield larger gains than inference-time injection, and whether similar strategies apply to other hallucination domains. The public release of MotionHalluc should accelerate research into more reliable multimodal reasoning systems.

• →Large multimodal AI models exhibit systematic hallucinations when comparing motion across paired videos across three dimensions: directional, attributional, and temporal errors.
• →Explicit kinematic measurement injection during inference improves motion reasoning accuracy by an average of 10.6% without requiring model retraining.
• →MotionHalluc benchmark provides 1,540 fine-grained evaluation questions across 553 video pairs for systematic assessment of motion reasoning reliability.
• →Quantitative measurement-grounded approaches may be essential for AI systems requiring precise physical reasoning beyond visual pattern recognition.
• →The training-free Perceive-Parse-Verify baseline offers immediate applicability to existing models, suggesting practical near-term solutions before deeper architectural changes.