Research

Exploring data use, representations, and generalization in generative models, with applications to visual learning and agent collaboration.

Understanding Model Collapse

Understanding Model Collapse

NeurIPS 2025 · Spotlight · Equal contribution

Explaining collapse as a transition from generalization to memorization, and studying entropy-based data selection for recursive training.

Paper · Project · Code

Diffusion modelsGeneralization & memorization
Representations & Generalization

Representations & Generalization

ICLR 2026 · Poster

Connecting balanced and spiky representations to generalization and memorization, with applications to detection and image editing.

Paper · Project · Code

Diffusion modelsGeneralization & memorization
Sim2Real for UAV Detection

Sim2Real for UAV Detection

CFHA · Preprint

Bridging synthetic and real UAV imagery through global style transfer, local refinement, and hallucination removal.

Paper · Code

Synthetic dataVisual learning
Multimodal Image Outpainting

Multimodal Image Outpainting

VIP · ACCV 2024

Combining multimodal descriptions and diffusion generation to customize the content beyond an image’s boundaries.

Paper · Code

Multimodal generationImage outpainting
Latent Communication Across Agents

Latent Communication Across Agents

See What I See · 2026 Preprint

Studying dense KV-cache alignment to transfer context and reasoning across different language models.

Paper · Project

Heterogeneous agentsLatent communication