An-Ya Ji

Synthesizing data environments from explicit world specifications, not just files, unlocks scalable training for data agents and reveals a fundamental computation-insight gap in frontier models.

Models trained on CoVA-SFT outperform existing multimodal reasoning baselines by over 2x, revealing the potential of structured visual-textual integration in AI.

OvisOCR2 outperforms traditional pipeline methods, achieving a groundbreaking score of 96.58 in end-to-end document parsing.

Models may give clearer instructions, but they fail to engage learners deeply, resulting in passive instruction-following rather than active understanding.

LLMs still struggle to go beyond simple lookups when answering questions about tables, especially when prediction and reasoning about unobserved data is required.