Physical & Multimodal Reasoning
Ongoing work on reasoning about the physical world (details coming soon)
This project studies how models can reason about the physical world rather than simply matching surface patterns. The goal is to move from perceptual fluency toward genuinely physically consistent understanding, so that a system can reason about how objects behave and interact over time and ground those judgments in what it actually observes. A further aim is to apply this kind of reasoning to generated visual content — assessing whether it is physically plausible and pointing toward ways it might be improved.
Focus
- Reasoning about physical dynamics, motion, and interactions as they unfold over time.
- Grounding reasoning in perception, so conclusions rest on observed evidence rather than assumptions.
- Assessing and improving the physical plausibility of generated content.
- Agentic, multi-step reasoning that decomposes a question and gathers the evidence needed to answer it.
This work is in progress; more details will be shared once it is published.