Vision Language Models Cannot Reason About Physical Transformation
This work identifies a critical limitation in current VLMs' ability to reason about physical transformations, which is a foundational problem for developing robust embodied AI agents.
This paper introduces ConservationBench, a benchmark to evaluate Vision Language Models' (VLMs) understanding of physical transformations and conservation of physical quantities. Across 23,040 questions and 112 VLMs, the study found that VLMs systematically fail, performing near chance on conservation tasks, with performance worsening when visual content is present despite strong textual priors for invariance.
Understanding physical transformations is fundamental for reasoning in dynamic environments. While Vision Language Models (VLMs) show promise in embodied applications, whether they genuinely understand physical transformations remains unclear. We introduce ConservationBench evaluating conservation -- whether physical quantities remain invariant under transformations. Spanning four properties with paired conserving/non-conserving scenarios, we generate 23,040 questions across 112 VLMs. Results reveal systematic failure: performance remains near chance with improvements on conservation tasks accompanied by drops on controls. Control experiments show strong textual priors favoring invariance, yet models perform worse with visual content. Neither temporal resolution, prompting, nor curated sampling helps. These findings show that current VLMs fail to maintain transformation-invariant representations of physical properties across dynamic scenes.