Rogério Feris

MLLMs struggle with multi-hop reasoning in document-style images, achieving only 62.83% accuracy compared to over 90% for human annotators.

Forget finetuning: TTA-Vid adapts video reasoning models to new datasets *during inference* using test-time reinforcement learning, achieving state-of-the-art results without any labels.

VLMs can now get a million-scale boost in chart-understanding abilities thanks to a new dataset with paired code, images, data, and reasoning.