DENG LAB · SJTU

Efficient AI,
across modalities.

School of Computer Science Shanghai Jiao Tong University

We develop methods and theory for efficient, effective, reliable, and trustworthy generative AI.

Our work spans unified multimodal models, world models and VLA, efficient architectures, and training, inference and RL infrastructure.

Our research Join the lab

DENG Lab members gathered around a conference table at Shanghai Jiao Tong University. — DENG Lab, Shanghai · 2025

Selected research

All publications

LatentUM overview showing self-reflective text-to-image generation, visual spatial planning, and world modeling using shared semantic visual tokens.

LatentUM

NeurIPS 2026Oral

A unified latent-space model enables interleaved reasoning across text and images.

Single-block and multi-block diffusion timelines comparing sequential decoding with overlapping block refinement and KV-cache storage.

MBD-LMs

EMNLP 2026Oral

Concurrent block decoding combines Multi-block Teacher Forcing with an efficient Block Buffer inference engine.

Paper Code Blog

CLLMs global consistency objective mapping intermediate states on a Jacobi decoding trajectory directly to the converged token sequence.

CLLMs

ICML 2024

Consistency training turns pretrained language models into fast parallel decoders without extra draft models.

Paper Code Blog

D2F overview: block-wise causal diffusion training aligned with a pretrained bidirectional diffusion model.

D2F

ICLR 2026

Discrete diffusion forcing enables KV caching and pipelined parallel decoding for faster diffusion language models.

Paper Code Blog

Orthus architecture with text and image embeddings, a shared autoregressive transformer, and language and diffusion output heads.

Orthus

ICML 2025

A shared autoregressive transformer generates interleaved text and images through modality-specific heads.

Mantis paper teaser showing its human-video, robot-demonstration and multimodal data sources, disentangled visual foresight, adaptive temporal ensemble, and embodied reasoning examples.

Mantis

CVPR 2026

A vision–language–action model uses disentangled visual foresight to connect prediction and action.

News

Sep 2026
LatentUM is accepted by NeurIPS 2026 as an Oral paper! 🎉
Aug 2026
MBD-LMs is accepted as an Oral paper at EMNLP 2026, and CForce is also accepted by EMNLP 2026! 🎉
Jun 2026
World-Language-Action Model is accepted by CoRL 2026! 🎉
Jun 2026
MirrorPPR is accepted by ECCV 2026! 🎉

Gallery

Explore Gallery

Streaming-WAM video cover: real-time world-action generation for robot manipulation.

Streaming-WAM

Watch on Xiaohongshu

TacForcing video cover: tactile feedback and streaming robot action generation.

TacForcing

Watch on Xiaohongshu

National Natural Science Foundation of China

Shanghai Municipal Science and Technology Commission

Sharpa

Kuaishou

Huawei

Tsinghua University

Zhipu AI

Alimama

Ant Group

Transsion

Shanghai Innovation Institute

Join our lab

Interested in fundamental AI research? We would love to hear from you.