Fucai Ke

Embodied VLMs routinely succeed for the wrong reasons, answering visual queries correctly despite completely missing the critical timestamps and viewpoints required to ground their answers.

Unearthing two distinct planning competencies in LLMs reveals that scaling up models enhances operational reasoning but leaves structural enumeration largely unchanged.

VLMs struggle to reason about visual scenes in adverse weather, losing significant segmentation accuracy as rain, snow, or fog intensifies.

Despite advances in vision-language models, reasoning across sparse, multi-view observations remains surprisingly unsolved, with current models barely outperforming random guessing on a new benchmark.