MindEdit-Bench: Benchmarking Object-Level Counterfactual Spatial Reasoning in VLMs from In-the-Wild Photos
arXiv preprint, 2026* Equal contribution.
Research
My work focuses on multimodal foundation models, embodied decision-making under real-time constraints, and the systems that make training and evaluation more efficient.
Publications & Projects
Current Questions
- Multimodal Foundation Models & World Action Models (WAMs): Exploring VLMs, UMMs, VLAs, and WAMs for perception, reasoning, world modeling, and action.
- Reinforcement Learning: Studying decision-making and policy learning in interactive environments.
- Efficient AI Systems: Building GPU/CUDA environment acceleration, KV streaming cache, FlexAttention masks, and training/inference optimizations.