We introduce FLAPS, a method for reconstructing functions and physical systems from sparse, noisy observations while quantifying uncertainty. Using pretrained generative models, it improves regression accuracy and solves scientific inverse problems with competitive or better accuracy at lower sampling cost. [NeurIPS 2026]

We show how to adapt molecular generative models to find promising candidates with fewer costly evaluations. Across six computational small-molecule design tasks, our approach matches the strongest inference-time search baseline using 58 times fewer evaluations. [NeurIPS 2026]

We introduce MorphMimic, a shared world model and motion-tracking policy that lets humanoid robots with different body designs learn together. Across three humanoids in simulation, it matches robot-specific specialists and completes more than 99.7% of motion clips on each robot; the shared policy also transfers directly to a physical Unitree G1 for motion tracking and teleoperation.

I wrote an opinion piece arguing that knowledge is emerging as a new scaling dimension for self-improving AI. Knowledge flywheels turn agent experience into reusable insights about what works, what fails, when, and why. This curated knowledge improves the agents, harnesses, and models that follow.

We introduce knowledge-centric self-improvement, a paradigm in which generic, stateless agents improve a shared, curated knowledge base rather than changing themselves. Across abstract reasoning, coding, and terminal tasks, this approach improves solve rates at lower cost and transfers across tasks and model families.

We introduce Vera, a layered diffusion framework for video editing that separates what to generate from what to preserve. Vera jointly produces an edit layer, an alpha matte, and a composite video, enabling more faithful content preservation while supporting creative edits.

We introduce SpeeDiff, an approach that makes high-quality image generation far more efficient to train. It learns the visual representation and generative model together, using pixel-level feedback to prevent representation collapse. SpeeDiff-XL achieves strong ImageNet results while training over 140x faster than Vanilla SiT and 61x faster than REPA. [CVPR 2026]

We introduce FormulaCode, a benchmark for evaluating how well coding agents can optimize large, real-world scientific software repositories. FormulaCode tests agents on realistic performance bottlenecks with expert-written fixes and community-maintained workloads, revealing where current agents still struggle with repository-scale optimization. [ICML 2026]