Causal Video & Interactive World Models
Diagnosing and reducing the representation drift that compounds across long autoregressive rollouts.
Research & projects
From long-horizon drift to low-step integration and irregular observations, each project starts with the system failure that matters.
Diagnosing and reducing the representation drift that compounds across long autoregressive rollouts.
Learning context-conditioned integration clocks for frozen generative ODE and flow-matching backbones.
Parallel generative forecasting over irregular timestamps using stable latent trajectories.
Building traceable agent foundations for restricted-domain financial and corporate-filing analysis.
History-conditioned flow matching for generating level-2 limit order books.
Adaptive Bayesian optimization for expensive, noisy portfolio-model tuning.
A source-first toolkit for transferring and evaluating fixed diffusion schedules in normalized flow time.
Estimating hidden width and propagation depth from a channel-capacity view of graph representation learning.
Decoupled graph diffusion for multi-stock movement prediction.
Multi-relational graph diffusion with parallel retention for stock-trend classification.