概率推断
Probabilistic inference
Graphical models, Bayesian inference, and scalable approximate inference algorithms.
AISTATS 2023 · Ising model selection ↗About PIL Lab
The Probabilistic Inference and Learning Lab is a research group at Zhejiang University. We study machine learning at the intersection of information theory, signal processing, and statistical physics.
Our work develops theory and algorithms for high-dimensional inference and generation. Current projects include diffusion and flow models, generative methods for inverse problems, and parallel decoding for diffusion language models.
Based at the ZJU–UIUC Institute ↗1 principal investigator · 10 graduate researchers · 1 visiting student
Research areas
Our research focuses on probabilistic inference, generative modeling, inverse problems, and diffusion language models.
生成建模
Generative modeling
Diffusion models, score-based generative models, flow matching, and their theoretical foundations.
Diffusion + Flow Reading Map ↗逆问题
Inverse problems
Compressed sensing, image restoration, and generative priors for inverse problems.
ACML 2024 · Diffusion posterior sampling ↗语言模型与智能体
Language models & agents
Diffusion language models, multimodal reasoning, parallel decoding, and agent methods.
NeurIPS 2026 · Parallel multimodal decoding ↗Selected publications
Selected work by the PI and lab members in diffusion language models, generative inverse problems, statistical learning, Bayesian inference, and signal processing.
NeurIPS
Visual-Redundancy-Controlled Parallel Decoding for Diffusion-Based Multimodal Large Language Models
A training-free decoder uses token-to-image attention to select visually complementary positions and reduce redundancy during parallel diffusion decoding.
Findings of EMNLP
Cluster-Level Attention-Guided Parallel Decoding for Masked Diffusion Language Models
Confidence-induced token clusters and self-attention dependencies enable conflict-aware span-level updates for faster masked diffusion decoding.
ACML · Best Paper Runner-Up
Diffusion Model Based Posterior Sampling for Noisy Linear Inverse Problems
IEEE Signal Processing Letters
A Unified Bayesian Inference Framework for Generalized Linear Models
Lab members
PIL Lab currently includes one principal investigator, four PhD students, six master's students, and one visiting student.
Principal Investigator
Xiangming Meng 孟祥明
Assistant Professor · PhD Supervisor
研究员、助理教授、博士生导师
ZJU–UIUC Institute, Zhejiang University
His research focuses on probabilistic generative modeling, diffusion and flow models, diffusion language models, and learning algorithms for high-dimensional inference and inverse problems.
PhD students 博士生
4 members
Master's students 硕士生
6 members
王棂煜
Master's Student
夏笑
Master's Student
涂宇轩
Master's Student
蓝郑挺
Master's Student
贾秀哲
Master's Student
孙天漫
Master's Student
Visiting student 访问学生
1 visitor
Ying Zhang张赢
Visiting Student
University of Cambridge
Reading maps
PIL Lab maintains two public reading maps for diffusion language models, diffusion models, and flow matching.
扩散语言模型阅读地图
dLLM Reading Map
A structured reading list on diffusion language models, covering foundations, methods, applications, and agents.
107 papers · 12-paper core path扩散模型与流匹配论文导航
Diffusion + Flow Reading Map
A structured reading list on diffusion models, score-based modeling, flow matching, and rectified flow.
110 papers · books, courses & tutorialsLab news
Publication
Steering Embedded Language Flows with Formal Constraints was accepted to DiffuLM @ NeurIPS 2026.
WebsiteThe first version of the PIL Lab website is online.
PreprintOur new preprint, One Latent, Many Tokens: Jointly Learning Compressed Embeddings for Efficient Language Diffusion, is now available on arXiv.
PublicationVisual-Redundancy-Controlled Parallel Decoding for Diffusion-Based Multimodal Large Language Models was accepted to NeurIPS 2026.
PublicationCluster-Level Attention-Guided Parallel Decoding for Masked Diffusion Language Models was accepted to Findings of EMNLP 2026.
ServiceXiangming Meng will serve as an Area Chair for ICLR 2027.
ServiceXiangming Meng will serve as an Area Chair for NeurIPS 2026.
PublicationUST-SSM: Unified Spatio-Temporal State Space Models for Point Cloud Video Modeling was accepted to ICCV 2025.
PublicationFIG: Flow with Interpolant Guidance for Linear Inverse Problems was accepted to ICLR 2025.
AwardThe DMPS paper received the ACML 2024 Best Paper Runner-Up Award.
For archived items, dates are based on the first public update of the PI homepage.
More updates on the PI homepage ↗Contact and opportunities
Students and researchers interested in probabilistic inference, generative models, diffusion language models, or inverse problems may contact the lab by email.