Two Neural Graph Database papers accepted: NGDB-Zoo (NeurIPS 2026) and NGDBench (EMNLP 2026 Findings). Congrats to Zhongwei and Yufei!
Invited guide talk at CCF YOCSEF HQ on code world models, covering PatchWorld and VisualPatchWorld.
VisualPatchWorld learns executable, inspectable code world models for visual control by combining active structure discovery, parameter fitting, and model-predictive planning.
We introduce PatchWorld, a gradient-free framework that induces inspectable, replayable, and locally repairable Python world models from offline agent trajectories for prediction and planning.
Enterprise agents need more than RAG. Our survey maps how LLMs can build the structured grounding layer—vocabulary, constraints, actions, governance, and maintenance—that makes agents actually deployable.
Led by Zhongwei Xie, our survey maps the four-layer Agent Harness stack—execution, context, interaction, and guardrails—and why it may matter as much as the model itself.
How LeCun's vision for world models and knowledge graph reasoning converge on the same architecture - and what that means for both fields.
A short overview of my Home Depot talk on building knowledge graphs, from rich semantics to autonomous construction and task-aware optimization.
