About

Volodymyr Kuleshov is an Assistant Professor in the Department of Computer Science at Cornell Tech and Cornell University. His research focuses on machine learning and its applications in science, health, and sustainability. His work involves two main directions: core research in machine learning, specifically generative models, probabilistic methods, diffusion-based language models, and decision-making under uncertainty; and the development of machine learning-based technologies that improve human and environmental health. His previous projects have focused on genome sequencing, machine reading, and reducing food waste. Kuleshov has been involved in commercializing his research on diffusion language models through the company Inception and co-founded Afresh, a startup that uses AI to significantly reduce food waste, which is now deployed in about 10% of US supermarkets. His earlier work on genome sequencing was commercialized by the Stanford spin-off Moleculo and became part of Illumina's genome phasing service. He obtained his PhD from Stanford University, where he received the Arthur Samuel Best Thesis Award and worked with notable researchers including Stefano Ermon, Serafim Batzoglou, Michael Snyder, Christopher Re, and Percy Liang.

Research topics

• Computer Science
• Artificial Intelligence
• Mathematics
• Theoretical computer science
• Algorithm
• Computer vision
• Arithmetic
• Mathematical optimization
• Engineering
• Data science
• Management science
• Knowledge management

Selected publications

• PlantCAD2: a DNA foundation model for interpreting genomes across flowering plants

  bioRxiv (Cold Spring Harbor Laboratory) · 2025-09-01 · 6 citations

  preprintOpen accessCorresponding

  Understanding how DNA sequence encodes biological function remains a fundamental challenge in biology. Flowering plants (angiosperms), the dominant terrestrial clade, exhibit maximal biochemical complexity, extraordinary species diversity (over 100,000 species), relatively recent origins (~160 million years), ~200-fold variation in genome size and relative compact coding regions compared with other eukaryotes. These features present both a unique challenge and opportunity for pre-training DNA language models to understand plant-specific evolutionary conservation, regulatory architectures and genomic functions. Here, we introduce PlantCAD2, a long-context, plant-specific DNA language model with single-nucleotide resolution, pre-trained on 65 angiosperm genomes, together with a series of public benchmarks for evaluation. Comprehensive zero-shot testing shows that PlantCAD2 (676 million parameters) efficiently captures evolutionary conservation, surpassing the 7-billion-parameter Evo2 model in 10 of 12 tasks. With parameter-efficient fine-tuning, PlantCAD2 also outperforms the 1-billion-parameter AgroNT across seven cross-species tasks including chromatin accessible region, gene expression and protein translation. Moreover, its 8,192bp context window substantially improves accessible chromatin prediction in large genomes such as maize (AUPRC increasing from 0.587 to 0.711), underscoring the importance of long-range context for modeling distal regulation. Together, these results establish PlantCAD2 as a powerful, efficient, and versatile foundation model for plant genomics, enabling accurate genome annotation across diverse species.

  PublisherOA PDFDOI

• RanDeS: Randomized Delta Superposition for Multi-Model Compression

  ArXiv.org · 2025-05-16

  preprintOpen access

  From a multi-model compression perspective, model merging enables memory-efficient serving of multiple models fine-tuned from the same base, but suffers from degraded performance due to interference among their task-specific parameter adjustments (i.e., deltas). In this paper, we reformulate model merging as a compress-and-retrieve scheme, revealing that the task interference arises from the summation of irrelevant deltas during model retrieval. To address this issue, we use random orthogonal transformations to decorrelate these vectors into self-cancellation. We show that this approach drastically reduces interference, improving performance across both vision and language tasks. Since these transformations are fully defined by random seeds, adding new models requires no extra memory. Further, their data- and model-agnostic nature enables easy addition or removal of models with minimal compute overhead, supporting efficient and flexible multi-model serving.

  PublisherOA PDFDOI

• d2: Improved Techniques for Training Reasoning Diffusion Language Models

  arXiv (Cornell University) · 2025-09-25

  preprintOpen accessSenior author

  While diffusion language models (DLMs) have achieved competitive performance in text generation, improving their reasoning ability with reinforcement learning remains an active research area. Here, we introduce d2, a reasoning framework tailored for masked DLMs. Central to our framework is a new policy gradient algorithm that relies on accurate estimates of the sampling trajectory likelihoods. Our likelihood estimator, d2-AnyOrder, achieves exact trajectory likelihood with a single model pass for DLMs that support a sampling algorithm called any-order decoding. Through an empirical study of widely used DLMs, we show that any-order decoding is not universally supported in practice. Consequently, for DLMs that do not naturally support any-order decoding, we propose another estimator, d2-StepMerge, which, unlike d2-AnyOrder, only approximates the trajectory likelihood. d2-StepMerge trades off compute for approximation accuracy in an analytically tractable manner. Empirically, d2 significantly outperforms widely-used RL baselines when applied to popular DLMs, and sets a new state-of-the-art performance for DLMs on logical reasoning tasks (Countdown and Sudoku) and math reasoning benchmarks (GSM8K and MATH500). We provide the code along with a blog post on the project page: https://guanghanwang.com/d2

  PublisherOA PDFDOI

• Probabilistic Graphical Models: A Concise Tutorial

  ArXiv.org · 2025-07-23

  preprintOpen accessSenior author

  Probabilistic graphical modeling is a branch of machine learning that uses probability distributions to describe the world, make predictions, and support decision-making under uncertainty. Underlying this modeling framework is an elegant body of theory that bridges two mathematical traditions: probability and graph theory. This framework provides compact yet expressive representations of joint probability distributions, yielding powerful generative models for probabilistic reasoning. This tutorial provides a concise introduction to the formalisms, methods, and applications of this modeling framework. After a review of basic probability and graph theory, we explore three dominant themes: (1) the representation of multivariate distributions in the intuitive visual language of graphs, (2) algorithms for learning model parameters and graphical structures from data, and (3) algorithms for inference, both exact and approximate.

  PublisherOA PDFDOI

• A foundational model for joint sequence-function multi-species modeling at scale for long-range genomic prediction

  bioRxiv (Cold Spring Harbor Laboratory) · 2025-12-25 · 9 citations

  articleOpen access

  Genomic prediction and design require models that integrate local sequence features with long-range regulatory dependencies spanning hundreds of kilobases to megabases. Existing approaches have made substantial progress along complementary axes: supervised sequence-to-function models achieve high accuracy for specific assays and organisms, self-supervised genomic foundation models learn transferable representations from large-scale sequence data, and conditional generative models enable principled sequence design guided by functional objectives. However, these strengths are typically realized in isolation - across distinct model classes, architectures, and training regimes - limiting the ability to combine long-context, base-resolution prediction, functional modeling, and controllable generation within a single efficient framework that generalizes across organisms and modalities. Here we introduce Nucleotide Transformer v3 (NTv3), a multi-species foundation model that unifies representation learning, functional-track and genome-annotation prediction, and controllable sequence generation within a common backbone. NTv3 uses a U-Net-like architecture to enable single-base tokenization and efficient modeling of contexts up to 1 Mb. We pre-train NTv3 on 9 trillion base pairs from OpenGenome2 using base-resolution masked language modeling, followed by post-training with a joint objective that integrates continued self-supervision with supervised learning on 16,000 functional tracks and annotation labels from 24 animal and plant species. After post-training, NTv3 achieves state-of-the-art accuracy for functional-track prediction and genome annotation across species, outperforming leading sequence-to-function and foundation-model baselines on established benchmarks and on the new NTv3 benchmark, a controlled downstream fine-tuning suite in a standardized 32 kb input / base-resolution output setting. We further show that NTv3 consolidates a shared regulatory grammar across tasks, enabling coherent long-range genome-to-function inference and variant-associated remodeling. Finally, we fine-tune NTv3 into a controllable generative model via masked diffusion language modeling and use it to design enhancer sequences with specified activity levels and promoter selectivity. We validate these designs experimentally, showing that generated enhancers recapitulate the intended activity stratification and achieve the desired promoter-specific activation in cellulo. We release the NTv3 model family together with code and practical cookbooks for long-context training, multispecies post-training, fine-tuning, interpretation, and sequence design.

  PublisherDOI

• Remasking Discrete Diffusion Models with Inference-Time Scaling

  arXiv (Cornell University) · 2025-03-01

  preprintOpen accessSenior author

  Part of the success of diffusion models stems from their ability to perform iterative refinement, i.e., repeatedly correcting outputs during generation. However, modern masked discrete diffusion lacks this capability: when a token is generated, it cannot be updated again, even when it introduces an error. Here, we address this limitation by introducing the remasking diffusion model (ReMDM) sampler, a method that can be applied to pretrained masked diffusion models in a principled way and that is derived from a discrete diffusion model with a custom remasking backward process. Most interestingly, ReMDM endows discrete diffusion with a form of inference-time compute scaling. By increasing the number of sampling steps, ReMDM generates natural language outputs that approach the quality of autoregressive models, whereas when the computation budget is limited, ReMDM better maintains quality. ReMDM also improves sample quality of masked diffusion models for discretized images, and in scientific domains such as molecule design, ReMDM facilitates diffusion guidance and pushes the Pareto frontier of controllability relative to classical masking and uniform noise diffusion. We provide the code along with a blog post on the project page: https://guanghanwang.com/remdm

  PublisherOA PDFDOI

• Simple Guidance Mechanisms for Discrete Diffusion Models.

  PubMed · 2025-04-01

  article

  Diffusion models for continuous data gained widespread adoption owing to their high quality generation and control mechanisms. However, controllable diffusion on discrete data faces challenges given that continuous guidance methods do not directly apply to discrete diffusion. Here, we provide a straightforward derivation of classifier-free and classifier-based guidance for discrete diffusion, as well as a new class of diffusion models that leverage uniform noise and that are more guidable because they can continuously edit their outputs. We improve the quality of these models with a novel continuous-time variational lower bound that yields state-of-the-art performance, especially in settings involving guidance or fast generation. Empirically, we demonstrate that our guidance mechanisms combined with uniform noise diffusion improve controllable generation relative to autoregressive and diffusion baselines on several discrete data domains, including genomic sequences, small molecule design, and discretized image generation. Code to reproduce our experiments is available here.

  Publisher

• Denoising Diffusion Variational Inference: Diffusion Models as Expressive Variational Posteriors

  Proceedings of the AAAI Conference on Artificial Intelligence · 2025-04-11 · 1 citations

  articleOpen accessSenior author

  We propose denoising diffusion variational inference (DDVI), a black-box variational inference algorithm for latent variable models which relies on diffusion models as flexible approximate posteriors. Specifically, our method introduces an expressive class of diffusion-based variational posteriors that perform iterative refinement in latent space; we train these posteriors with a novel regularized evidence lower bound (ELBO) on the marginal likelihood inspired by the wake-sleep algorithm. Our method is easy to implement (it fits a regularized extension of the ELBO), is compatible with black-box variational inference, and outperforms alternative classes of approximate posteriors based on normalizing flows or adversarial networks. We find that DDVI improves inference and learning in deep latent variable models across common benchmarks as well as on a motivating task in biology-inferring latent ancestry from human genomes-where it outperforms strong baselines on 1000 Genomes dataset.

  PublisherOA PDFDOI

• Calibrated Probabilistic Forecasts for Arbitrary Sequences.

  PubMed · 2025-03-01

  article

  Real-world data streams can change unpredictably due to distribution shifts, feedback loops and adversarial actors, which challenges the validity of forecasts. We present a forecasting framework ensuring valid uncertainty estimates regardless of how data evolves. Leveraging the concept of Blackwell approachability from game theory, we introduce a forecasting framework that guarantees calibrated uncertainties for outcomes in any compact space (e.g., classification or bounded regression). We extend this framework to recalibrate existing forecasters, guaranteeing calibration without sacrificing predictive performance. We implement both general-purpose gradient-based algorithms and algorithms optimized for popular special cases of our framework. Empirically, our algorithms improve calibration and downstream decision-making for energy systems.

  Publisher

• The GAN is dead; long live the GAN! A Modern GAN Baseline

  arXiv (Cornell University) · 2025-01-09 · 4 citations

  preprintOpen access

  There is a widely-spread claim that GANs are difficult to train, and GAN architectures in the literature are littered with empirical tricks. We provide evidence against this claim and build a modern GAN baseline in a more principled manner. First, we derive a well-behaved regularized relativistic GAN loss that addresses issues of mode dropping and non-convergence that were previously tackled via a bag of ad-hoc tricks. We analyze our loss mathematically and prove that it admits local convergence guarantees, unlike most existing relativistic losses. Second, our new loss allows us to discard all ad-hoc tricks and replace outdated backbones used in common GANs with modern architectures. Using StyleGAN2 as an example, we present a roadmap of simplification and modernization that results in a new minimalist baseline -- R3GAN. Despite being simple, our approach surpasses StyleGAN2 on FFHQ, ImageNet, CIFAR, and Stacked MNIST datasets, and compares favorably against state-of-the-art GANs and diffusion models.

  PublisherOA PDFDOI

Frequent coauthors

• Stefano Ermon

  20 shared

• Aaron Gokaslan

  8 shared

• M Snyder

  7 shared

• Alexander M. Rush

  6 shared

• Serafim Batzoglou

  6 shared

• Shachi Deshpande

  Cornell University

  6 shared

• Yair Schiff

  6 shared

• Fei Wang

  Boehringer Ingelheim (United States)

  4 shared

Labs

• Volodymyr Kuleshov LabPI

  Machine learning and its applications in science, health, and sustainability

Education

• Ph.D., Computer Science

  Stanford University

Awards & honors

• Arthur Samuel Best Thesis Award
• NSF CAREER Award
• NIH MIRA Award