Published in ICCV 2025 Oral, 2025
Learns cognitively aligned layer-by-layer SVG construction with a diffusion transformer, producing editable vector graphics with meaningful semantic structure.
Published in CCF-A Conference Under Review, 2026
Separates low-level player control from high-level NPC strategy, enabling controllable interactions and zero-shot strategy transfer across game world models.
Published in CoRL 2026, 2026
Introduces geometry-guided representation alignment for adapting vision-language-action models from synthetic robot videos while keeping low-level control grounded in real demonstrations.
Published in CCF-A Conference Under Review, 2026
Decouples state understanding, NPC decision-making, action control, and video generation, then reconnects them in a closed interaction loop for state-aware NPC behavior.
Published in Technical Report, 2026
Turns the language understanding already learned by a large video generator into precise, temporally grounded character and camera control with lightweight adaptation.