PUBLICATIONS & PREPRINTS

Research.

Work on visual generation, diffusion models, and language models.

CRP: a history-free base prediction plus a history-conditioned residual, using one network in two passes.

arXiv 2026Video generation · CRP

Conditional Residual Prediction: Improving Autoregressive Video Diffusion without a Bidirectional Teacher

Bowen Zheng, Zhiguang Liu, Jiarong Ou, Rui Chen, Tianyang Hu

Reducing reliance on generated history through conditional residual prediction, enabling autoregressive video diffusion without a bidirectional teacher.

Read on arXiv

Prologue: separate prologue and visual tokens during training and autoregressive inference.

Core idea ↗

arXiv 2026Visual generation · Prologue

Autoregressive Visual Generation Needs a Prologue

Bowen Zheng, Weijian Luo, Guang Yang, Colin Zhang, Tianyang Hu

Learning a separate sequence of prologue tokens for autoregressive generation while keeping visual tokens dedicated to reconstruction.

Read on arXiv Code ↗

VCQ: comparing constant and variable codebook sizes and their effect on remaining entropy.

Core idea ↗

arXiv 2026Visual tokenization · VCQ

Taming the Entropy Cliff: Variable Codebook Size Quantization for Autoregressive Visual Generation

Bowen Zheng, Weijian Luo, Guang Yang, Colin Zhang, Tianyang Hu

Varying codebook capacity across token positions to address the entropy cliff in autoregressive visual generation.

Read on arXiv

wAR-Tok: aligning the tokenizer distribution with an autoregressive prior through Wasserstein gradient flow.

Core idea ↗

arXiv 2026Autoregressive priors · wAR-Tok

Learning Discrete Autoregressive Priors with Wasserstein Gradient Flow

Bowen Zheng, Yihong Luo, Tianyang Hu

Aligning discrete image tokens with an autoregressive prior through a Wasserstein-gradient-flow update during tokenizer training.

Read on arXiv

D2O: generative pre-training followed by GAN fine-tuning for one-step image generation.

Core idea ↗

ICML 2025One-step generation

Revisiting Diffusion Models: From Generative Pre-training to One-Step Generation

Bowen Zheng, Tianming Yang

Viewing diffusion training as generative pre-training, with lightweight GAN fine-tuning to unlock one-step generation.

Read on arXiv

SimLens: intermediate-layer predictions and an entropy threshold for early exit.

Core idea ↗

EMNLP 2026Efficient inference

SimLens for Early Exit in Large Language Models: Eliciting Accurate Latent Predictions with One More Token

Ming Ma*, Bowen Zheng*, Zhongqiao Lin, Tianming Yang

* Co-first authors

Decoding intermediate-layer predictions with a short continuation to support efficient early-exit inference.

Read on arXiv

Label Words: demonstrations of knowledge and categorization tasks in in-context learning.

Core idea ↗

Findings of ACL 2026In-context learning

Label Words as Local Task Vectors in In-Context Learning

Bowen Zheng*, Ming Ma*, Zhongqiao Lin, Tianming Yang

* Co-first authors

Studying how individual demonstrations form local task vectors and how language models aggregate their information.

Read on arXiv

Conditional Residual Prediction: Improving Autoregressive Video Diffusion without a Bidirectional Teacher ↗

Autoregressive Visual Generation Needs a Prologue ↗

Taming the Entropy Cliff: Variable Codebook Size Quantization for Autoregressive Visual Generation ↗

Learning Discrete Autoregressive Priors with Wasserstein Gradient Flow ↗

Revisiting Diffusion Models: From Generative Pre-training to One-Step Generation ↗

SimLens for Early Exit in Large Language Models: Eliciting Accurate Latent Predictions with One More Token ↗

Label Words as Local Task Vectors in In-Context Learning ↗

Conditional Residual Prediction: Improving Autoregressive Video Diffusion without a Bidirectional Teacher

Autoregressive Visual Generation Needs a Prologue

Taming the Entropy Cliff: Variable Codebook Size Quantization for Autoregressive Visual Generation

Learning Discrete Autoregressive Priors with Wasserstein Gradient Flow

Revisiting Diffusion Models: From Generative Pre-training to One-Step Generation

SimLens for Early Exit in Large Language Models: Eliciting Accurate Latent Predictions with One More Token

Label Words as Local Task Vectors in In-Context Learning