|
Research
My research develops systems, methods, and applications for
agentic reasoning, learning, and continual improvement,
with a focus on:
-
World and agent simulation:
Open-ended environment and agent simulation for
long-horizon reasoning, learning through interaction, and environment-agent self-improvement.
-
SimWorld (2025):
The new simulation infrastructure designed around coding agents
for building physical/social worlds and agent simulations, enabling
agent-environment recursive self-improvement, (embodied/social/multi-agent...) reasoning,
and safety evaluation.
-
Pandora (2024)
and
PAN (2025):
The first general video-language interactive world models for
simulative reasoning.
-
Reasoning and (reinforcement) learning:
Algorithms and systems for complex reasoning,
including planning with world models (e.g., RAP, 2023), new paradigm of continuous latent-space
reasoning (e.g., Coconut, 2025; PAN, 2025), efficient reinforcement learning (e.g., IsoCompute Playbook, 2026; Q learning for language, 2021), and unified learning mechanism from all experience (e.g., "Standard Model", 2021).
-
AI for Science:
Agents connecting scientific reasoning with data analysis and
simulation, including, e.g., single-cell biology and automated density
functional theory (DFT) workflows for materials science.
(scPilot, 2025;
TritonDFT, 2026)
|