Yingnian Wu

University of California,
Los Angeles

We propose a novel class of language models, Latent Thought Models (LTMs), which incorporate explicit latent thought vectors that follow an explicit prior model in latent space. These latent thought vectors guide the autoregressive generation of ground tokens through a Transformer decoder. Training employs a dual-rate optimization process within the classical variational Bayes framework. Empirical studies reveal that LTMs possess additional scaling dimensions beyond traditional Large Language Models (LLMs), such as the number of iterations in inference-time computation and number of latent thought vectors. Higher sample efficiency can be achieved by increasing training compute per token, with further gains possible by trading model size for more inference steps. Designed based on these scaling properties, LTMs demonstrate superior sample and parameter efficiency compared to autoregressive models and discrete diffusion models. They significantly outperform these counterparts in validation perplexity and zero-shot language modeling tasks. Additionally, LTMs exhibit emergent few-shot in-context reasoning capabilities that scale with model size, and achieve competitive performance in conditional and unconditional text generation.

The project page is available at https://deqiankong.github.io/blogs/ltm. This is a collaboration between UCLA and Lambda.

Shane Jensen

University of Pennsylvania

I will discuss Jun Liu's substantial impact on Bayesian nonparametrics and clustering through the widespread use of the Dirichlet process. I will present two applications of the Dirichlet process (DP), starting with early work on clustering on the micro level of DNA sequence data in order to find conserved patterns in genomic sequences. I will then discuss more recent work on DP-based clustering of crime at the macro level of Philadelphia neighborhoods. Finally, I will bring the micro and macro closer together in an extension of our DP-based approach to clustering crime across multiple levels of resolution simultaneously.

Xinran Li

University of Chicago

Introducing auxiliary data or parameters can significantly facilitate Bayesian computation, while treating known parameters as unknown can, in some cases, improve Frequentist inference. In this talk, I will present a bootstrap approach inspired by these ideas for conducting inference on partially identified parameters in the presence of missing data. Specifically, we focus on parameters defined by a set of over-identified estimating equations involving unobserved or missing variables subject to general constraints. An important application and the main motivation for this framework is sensitivity analysis for unmeasured confounding in observational studies.

In contrast to settings without missing data, we choose not to reduce the number of estimating equations by combining them, since each equation may provide useful information for constraining the unobserved variables. However, over-identification, meaning there are more equations than parameters, creates challenges for existing bootstrap methods, which may fail to produce valid confidence sets. To address this issue, we propose a bootstrap procedure that introduces auxiliary data and parameters. While these auxiliary components can be conceptual and do not affect the actual implementation, they play a critical role in establishing the validity of the proposed inference approach.

Ke Deng

Tsinghua University

The great success of large deep learning models in various applications in recent years have encouraged many researchers to seek improved performance by utilizing larger models and bigger data in practical problems involving unstructured data, leading to psychological implications to pursuit large models everywhere. However, the fundamental principle of statistical modelling tells us that an over-flexible large model without a clear focus on unique features of the problem of interest would often lead to inefficient utilization of data and sub-optimal results. In this talk, I will provide a few examples in analyzing complex unstructured data, including texts, videos and multimodal data, that deep learning can be greatly enhanced by statical learning once we integrate them wisely. We hope these examples could inspire more research efforts on developing advanced statistical approaches with competitive performance and transparent interpretation for analyzing complex unstructured data on top of deep learning.

Hongyu Zhao

Yale University

Understanding cellular responses to genetic perturbations is essential for deciphering gene regulation and phenotype formation. While high-throughput single-cell RNA-sequencing has facilitated detailed profiling of heterogeneous transcriptional responses to perturbations at the single-cell level, there remains a pressing need for computational models that can decode the mechanisms driving these responses and accurately predict outcomes to prioritize target genes for experimental design. We will introduce a deep generative learning framework designed to model and predict single-cell transcriptional responses to genetic perturbations, including single-gene and combinatorial multi-gene perturbations. Our method can effectively integrate prior biological knowledge and disentangle basal cell states from perturbation-specific salient representations by leveraging gene embeddings derived from large language models. Through comprehensive evaluations on multiple single-cell CRISPR Perturb-seq datasets, our method outperformed state-of-the-art methods in predicting perturbation outcomes, achieving higher prediction accuracy. Notably, it demonstrated robust generalization to unseen target genes and perturbations, and its predictions captured both average expression changes and the heterogeneity of single-cell responses. Furthermore, its predictions enable diverse downstream analyses, including identifying differentially expressed genes and exploring genetic interactions, demonstrating its utility and versatility. This is joint work with Gefei Wang, Tianyu Liu, Jia Zhao, and Youshu Cheng.

Chiara Sabatti

Stanford University

A common characteristics of genomic studies is that they measure a large number of variables. And it is typically of interest to learn which, among these many, is relevant for an outcome. What does it mean for a variable to be "important"? How can we capture the fact that the relevance of explanatory variables might change across different contexts? I will describe some recent progress in this area.

Shirley Liu

GV20 Therapeutics /
Dana-Farber Cancer Institute

GV20 Therapeutics is a clinical-stage, AI-driven biotherapeutics company focused on next-generation cancer treatments. The company’s lead program in development as a potential cancer immunotherapy is GV20-0251—a monoclonal antibody targeting the novel immune checkpoint IGSF8. It is the first AI-designed antibody against an AI-predicted target to enter clinical trials for treating advanced solid tumors. GV20’s proprietary STEAD AI platform enabled the program to advance from target discovery to investigational new drug (IND) approval in just three years, with early clinical data showing a favorable safety profile and promising efficacy in patients with advanced metastatic cancer.

Haiyan Huang

University of California,
Berkeley

I will share a line of my research that led me from pure mathematics into genomic data analysis, navigating challenges at the intersection of statistics, biology, and data science. I will also reflect on how Jun has influenced and inspired me throughout this journey.

Leo Guo

Bayesian Investment

In these brief remarks, I hope to share a few personal reflections on what an honor it is to celebrate Professor Jun Liu's two incredible milestones. I'll begin with the surreal and often humorous feeling of returning to campus after twelve years, recalling some of the shared struggles—like agonizing over a stubborn statistical model—that shaped us in his lab. From there, I want to focus on two profound gifts I received from his mentorship: the invaluable lesson in resilience, and the powerful philosophy that statistics is much more than numbers, but a way of seeing the world. I'll briefly touch on how these principles became my compass long after I left academia, guiding me through the challenges of the startup world. Ultimately, this is simply my heartfelt way of saying thank you, and of celebrating the incredible, lasting impact of a great mentor.

Jinfeng Zhang

Insilicom LLC

To address the rapid growth of scientific publications and data in biomedical research, knowledge graphs (KGs) have become a critical tool for integrating large volumes of heterogeneous data to enable efficient information retrieval and automated knowledge discovery. However, transforming unstructured scientific literature into KGs remains a significant challenge, with previous methods unable to achieve human-level accuracy. Here we used an information extraction pipeline that won first place in the LitCoin Natural Language Processing Challenge (2022) to construct a large-scale KG named iKraph using all PubMed abstracts. The extracted information matches human expert annotations and significantly exceeds the content of manually curated public databases. To enhance the KG’s comprehensiveness, we integrated relation data from 40 public databases and relation information inferred from high-throughput genomics data. This KG facilitates rigorous performance evaluation of automated knowledge discovery, which was infeasible in previous studies. We designed an interpretable, probabilistic-based inference method to identify indirect causal relations and applied it to real-time COVID-19 drug repurposing from March 2020 to May 2023. Our method identified around 1,200 candidate drugs in the first 4 months, with one-third of those discovered in the first 2 months later supported by clinical trials or PubMed publications. These outcomes are very challenging to attain through alternative approaches that lack a thorough understanding of the existing literature. A cloud-based platform (https://biokde.insilicom.com) was developed for academic users to access this rich structured data and associated tools.

Huimin Cheng

Boston University

As spatially resolved transcriptomics (SRT) datasets increasingly span multiple adjacent or replicated slices, effective joint analysis across slices is needed to reconstruct tissue structures and identify consistent spatial gene expression patterns. This requires resolving spatial correspondences between slices while capturing shared transcriptomic features, two tasks that are typically addressed in isolation. Multi-slice analysis remains challenging due to physical distortions, technical variability, and batch effects. To address these challenges, we introduce Joint Alignment and Deep Embedding for multi-slice SRT (JADE), a unified computational framework that simultaneously learns spatial location-wise alignments and shared low-dimensional embeddings across tissue slices. Unlike existing methods, JADE adopts a roundtrip framework in which each iteration alternates between alignment and embedding refinement. To infer alignment, we employ attention mechanisms that dynamically assess and weight the importance of different embedding dimensions, allowing the model to focus on the most alignment-relevant features while suppressing noise. To the best of our knowledge, JADE is the first method that jointly optimizes alignment and representation learning in a shared latent space, enabling robust multi-slice integration. We demonstrate that JADE outperforms existing alignment and embedding methods across multiple evaluation metrics in the 10x Visium human dorsolateral prefrontal cortex (DLPFC) and Stereo-seq axolotl brain datasets. By bridging spatial alignment and feature integration, JADE provides a scalable and accurate solution for cross-slice analysis of SRT data.

Steve Qin

Emory University

Bayesian hierarchical model is a powerful tool that enables borrowing strength among exchangeable experiments. However, inference results are not reliable if the exchangeable condition is compromised. We proposed some adaptive strategy to overcome this challenge. We showed its applications in analyzing high-throughput genomics data.

Xin Xing

Virginia Tech

In this talk, I will trace the remarkable evolution of statistical inference from its classical foundations to cutting-edge generative approaches. We begin with the classical p-value framework that established the theoretical bedrock of 20th-century statistics, then explore how mirror statistics and data-splitting methods emerged to address the multiple testing crisis in high-dimensional settings by cleverly exploiting distributional symmetries. The talk culminates with modern flow-based predictive inference methods that revolutionize uncertainty quantification by learning flexible data representations, enabling robust inference even when traditional exchangeability assumptions fail and testing data is contaminated. I will show how these three parts complement each other in practice, forming a unified toolkit that empowers practitioners to tackle diverse inference challenges in high-dimensional and complex data.

Junni Zhang

Peking University

Abstract: There is a growing demand for disaggregated demographic estimates to support both analysis and policymaking, driving the use of increasingly available, highly detailed demographic data. However, such data are often sparse or noisy. To address this, we combine mathematical demography with Bayesian statistics to leverage similar—but not identical—demographic patterns and to effectively handle missing data. We have developed a fast and scalable R package that implements a general family of Bayesian models, making these methods both accessible and practical for demographers.

Yuguo Chen

University of Illinois Urbana-Champaign

Sequential importance sampling (SIS) is a versatile and powerful Monte Carlo method for tackling complex statistical inference problems. It has had a profound impact on modern statistical computation, thanks to Professor Jun Liu’s pioneering contributions. In this talk, I will share a few projects related to SIS that I was fortunate to work on with Professor Liu and my students.

Roee Gutman

Brown University

Linking data from various sources is essential for generating accurate statistics and assessing the impact of interventions. With declining survey response rates, rapid technological advancements, and the growing volume of data collected by different organizations the demand for robust record linkage methods continues to grow. A major challenge in the linkage process arises when unique identifying information is unavailable across datasets, because of privacy concerns or errors in recording. To address this challenge there is a need for efficient computational algorithms and for statistical inference tools that adjust for false matches and missed links. We view record linkage as a missing data problem and develop Bayesian procedures that enhance the linkage results while ensuring valid statistical inferences.

Mayetri Gupta

University of Glasgow

Advances in genomic sequencing technologies in the past few decades have opened up the possibility of making previously unthinkable biological discoveries at extremely high resolution- but have led to numerous challenging problems in how to sensibly and accurately analyse the generated data. These data are typically of huge dimension and/or large volumes; may be discrete in nature; are subject to various artefacts; and their distributions exhibit non-standard features, such as long-ranging correlations, non-ellipsoidal shapes, skewness and multimodality- which may cause difficulties in making successful inference through usual statistical modelling approaches. In this lightning talk, I will highlight some recent examples of work from our group where detail-oriented Bayesian modelling coupled with robust, efficient and powerful Monte Carlo-based computational methods enabled useful inference and meaningful biological discoveries. These include (i) clustering high volumes of non-ellipsoidal data from genotype sequencing experiments; (ii) detecting differential epigenetic profiles from high-throughput DNA methylation data and (iii) discovering groups of genetic variants associated with complex phenotypic traits in humans. This talk is based on joint work with Edoardo Redivo, Hien D. Nguyen, Huizi Zhang, Ben Swallow, Tushar Ghosh, Lanxin Li, Vincent Macaulay, Peter D. Adams, members of the Framingham Bone research consortium and the BHF Institute of Cardiovascular & Medical Sciences at Glasgow.

Scott Schmidler

Duke University

A major challenge in vaccine design is the elicitation of broadly-neutralizing antibodies (bnAbs) that protect against viral escape. Stochastic models of affinity maturation play a critical role in understanding barriers to success and guiding novel design strategies. A key difficulty is phylogenetic inference in the presence of site-dependent evolutionary processes, a long-standing challenge in computational biology. We describe a sequential Monte Carlo algorithm for approximating marginal sequence likelihoods under site-dependence and provide finite sample bounds on the approximation error. We extend this to a combined data-augmentation and importance sampling scheme for approximating posterior probabilities of phylogenetic trees and reconstructing ancestral sequences under site-dependent models of molecular evolution. We apply this approach to reconstruct maturation lineages from high-throughput B cell receptor repertoire sequencing data, and demonstrate the impact of accounting for context-dependence on reconstruction accuracy. We then apply our models to the problem of choosing targets for the design of boosting immunogens aimed at elicitation of bnAbs for HIV.

Augustine Kong

University of Oxford

Tony Cai

University of Pennsylvania

In this talk, I will present results on minimax and adaptive transfer learning for nonparametric classification under a posterior drift model with distributed differential privacy constraints. We first establish the minimax misclassification rate, characterizing how privacy constraints and both source and target samples influence classification accuracy. The results reveal phase transition phenomena and highlight fundamental trade-offs between privacy and accuracy. Building on this, we propose a data-driven adaptive classifier that achieves the optimal rate (up to a logarithmic factor) over a broad class of parameter spaces, while respecting the same privacy constraints.

Edoardo Airoldi

Temple University

High-throughput technology (eg, sequencing, mass spec) allows us to quantify biological mechanisms at a resolution that array technology and small scale experiments cannot. Currently, a substantial portion of biological research is leveraging some of these technologies. This flexibility comes with a price, however. Modern high-throughput instrumentation relies on built-in data collection protocols that are often biased. (For instance, a mass spec selects the most abundant ions, at an early stage of the measurement process, for further analysis.) The major unexpected consequence of such protocols is that they carry information about those quantities we are interested in estimating (absolute protein abundance, in the mass spec example). Scientists that do not account for this information during analysis, whether by counting or estimation using a statistical model, will likely base their scientific conclusions on misleading numbers, even in simple experimental conditions. This statistical issue is poorly understood by practitioners and amateur statisticians alike. It is arguably the main challenge we need to tackle to produce valid scientific conclusions in the era of high-throughput technology. I'll provide an illustration in mass spectrometry.

Qing Zhou

University of California,
Los Angeles

Learning causal structures from observational data is a central challenge in statistics and machine learning. A common assumption in causal discovery is that data samples are i.i.d., yet this assumption is often violated in real-world settings such as social networks, brain imaging, and cell differentiation, where dependencies exist across observational units. In this talk, I present a unified framework for causal structure learning from dependent data through decorrelation. For continuous data, we extend the Gaussian DAG model by incorporating sample dependence over an undirected graph. For binary data, we introduce a latent utility model with correlated errors across units, under a probit regression-based DAG formulation. In both cases, we develop methods to estimate the between-unit covariance structure and use it to decorrelate the data, enabling the application of standard DAG learning algorithms. Numerical experiments on both simulated and real datasets show that these decorrelation methods significantly improve the accuracy of causal graph recovery compared to directly applying standard methods to dependent data.

Xihong Lin

Harvard University