Chunhua Shen

Open-AoE transforms egocentric video capture into a powerful resource for embodied intelligence, making it easier than ever to train robots with human-like manipulation skills.

Transforming a single image into a robust simulation environment could revolutionize how robots learn and evaluate policies in real-world scenarios.

A unified definition of world models could catalyze breakthroughs across AI subfields by clarifying what these internal simulators should predict and how they should be constructed.

MemDreamer narrows the performance gap with human experts in long video understanding to just 3.7 points while processing only 2% of the full context.

Current MLLMs struggle with fine-grained spatial reasoning, achieving only 37.2 F1 on challenging tasks compared to human performance of 84.0 F1.

Foundation models struggle with spatial tasks, achieving only 12% success in reproducing target viewpoints, but a novel post-training framework boosts performance to over 51%.

Achieve superior camera control and visual fidelity in video re-rendering by training on unpaired real and synthetic data with a novel metric geometry reward.

OmniJigsaw reveals a "bi-modal shortcut phenomenon" in joint audio-visual integration, demonstrating that naive fusion can be surprisingly ineffective and highlighting the importance of carefully designed cross-modal training strategies.

Doubling the number of tokens in a ViT-based autoencoder, combined with staged compression and self-supervised pretraining, dramatically improves generative performance under deep compression, without increasing the latent budget.

Diffusion language models can now efficiently self-evaluate their output quality by regenerating their own sequences, enabling more reliable uncertainty quantification and flexible-length generation.