Yuming Yang

Even state-of-the-art Video-LLMs fail to generalize to dynamic aerial environments once evaluation moves beyond repetitive question templates and static satellite imagery.

Low-probability tokens disproportionately influence model updates, and a simple reweighting strategy can significantly enhance performance without sacrificing generalization.

Existing models miss critical visual evidence in metaphor understanding, but M$^3$R-Reasoner closes the gap, outperforming larger models in both accuracy and justification metrics.