Volodymyr Kuleshov

John X. Morris, Volodymyr Kuleshov, Vitaly Shmatikov et al.

How much private information do text embeddings reveal about the original text? We investigate the problem of embedding \textit{inversion}, reconstructing the full text represented in dense text embeddings. We frame the problem as controlled generation: generating text that, when reembedded, is close to a fixed point in latent space. We find that although a naïve model conditioned on the embedding performs poorly, a multi-step method that iteratively corrects and re-embeds text is able to recover $92\%$ of $32\text{-token}$ text inputs exactly. We train our model to decode text embeddings from two state-of-the-art embedding models, and also show that our model can recover important personal information (full names) from a dataset of clinical notes. Our code is available on Github: \href{https://github.com/jxmorris12/vec2text}{github.com/jxmorris12/vec2text}.

Jerry Chee, Yaohui Cai, Volodymyr Kuleshov et al.

This work studies post-training parameter quantization in large language models (LLMs). We introduce quantization with incoherence processing (QuIP), a new method based on the insight that quantization benefits from $\textit{incoherent}$ weight and Hessian matrices, i.e., from the weights being even in magnitude and the directions in which it is important to round them accurately being unaligned with the coordinate axes. QuIP consists of two steps: (1) an adaptive rounding procedure minimizing a quadratic proxy objective; (2) efficient pre- and post-processing that ensures weight and Hessian incoherence via multiplication by random orthogonal matrices. We complement QuIP with the first theoretical analysis for an LLM-scale quantization algorithm, and show that our theory also applies to an existing method, OPTQ. Empirically, we find that our incoherence preprocessing improves several existing quantization algorithms and yields the first LLM quantization methods that produce viable results using only two bits per weight. Our code can be found at https://github.com/Cornell-RelaxML/QuIP.

Yuntian Deng, Volodymyr Kuleshov, Alexander M. Rush · allen-ai

Language models have demonstrated the ability to generate highly fluent text; however, it remains unclear whether their output retains coherent high-level structure (e.g., story progression). Here, we propose to apply a statistical tool, model criticism in latent space, to evaluate the high-level structure of the generated text. Model criticism compares the distributions between real and generated data in a latent space obtained according to an assumptive generative process. Different generative processes identify specific failure modes of the underlying model. We perform experiments on three representative aspects of high-level discourse -- coherence, coreference, and topicality -- and find that transformer-based language models are able to capture topical structures but have a harder time maintaining structural coherence or modeling coreference.

Aaron Gokaslan, A. Feder Cooper, Jasmine Collins et al.

We assemble a dataset of Creative-Commons-licensed (CC) images, which we use to train a set of open diffusion models that are qualitatively competitive with Stable Diffusion 2 (SD2). This task presents two challenges: (1) high-resolution CC images lack the captions necessary to train text-to-image generative models; (2) CC images are relatively scarce. In turn, to address these challenges, we use an intuitive transfer learning technique to produce a set of high-quality synthetic captions paired with curated CC images. We then develop a data- and compute-efficient training recipe that requires as little as 3% of the LAION-2B data needed to train existing SD2 models, but obtains comparable quality. These results indicate that we have a sufficient number of CC images (~70 million) for training high-quality models. Our training recipe also implements a variety of optimizations that achieve ~3X training speed-ups, enabling rapid model iteration. We leverage this recipe to train several high-quality text-to-image models, which we dub the CommonCanvas family. Our largest model achieves comparable performance to SD2 on a human evaluation, despite being trained on our CC dataset that is significantly smaller than LAION and using synthetic captions for training. We release our models, data, and code at https://github.com/mosaicml/diffusion/blob/main/assets/common-canvas.md

Yingheng Wang, Yair Schiff, Aaron Gokaslan et al.

While diffusion models excel at generating high-quality samples, their latent variables typically lack semantic meaning and are not suitable for representation learning. Here, we propose InfoDiffusion, an algorithm that augments diffusion models with low-dimensional latent variables that capture high-level factors of variation in the data. InfoDiffusion relies on a learning objective regularized with the mutual information between observed and hidden variables, which improves latent space quality and prevents the latents from being ignored by expressive diffusion-based decoders. Empirically, we find that InfoDiffusion learns disentangled and human-interpretable latent representations that are competitive with state-of-the-art generative and contrastive methods, while retaining the high sample quality of diffusion models. Our method enables manipulating the attributes of generated images and has the potential to assist tasks that require exploring a learned latent space to generate quality samples, e.g., generative design.

Richa Rastogi, Yair Schiff, Alon Hacohen et al. · allen-ai

We introduce semi-parametric inducing point networks (SPIN), a general-purpose architecture that can query the training set at inference time in a compute-efficient manner. Semi-parametric architectures are typically more compact than parametric models, but their computational complexity is often quadratic. In contrast, SPIN attains linear complexity via a cross-attention mechanism between datapoints inspired by inducing point methods. Querying large training sets can be particularly useful in meta-learning, as it unlocks additional training signal, but often exceeds the scaling limits of existing models. We use SPIN as the basis of the Inducing Point Neural Process, a probabilistic model which supports large contexts in meta-learning and achieves high accuracy where existing models fail. In our experiments, SPIN reduces memory requirements, improves accuracy across a range of meta-learning tasks, and improves state-of-the-art performance on an important practical problem, genotype imputation.

Shachi Deshpande, Kaiwen Wang, Dhruv Sreenivas et al. · microsoft-research

Estimating the effect of intervention from observational data while accounting for confounding variables is a key task in causal inference. Oftentimes, the confounders are unobserved, but we have access to large amounts of additional unstructured data (images, text) that contain valuable proxy signal about the missing confounders. This paper argues that leveraging this unstructured data can greatly improve the accuracy of causal effect estimation. Specifically, we introduce deep multi-modal structural equations, a generative model for causal effect estimation in which confounders are latent variables and unstructured data are proxy variables. This model supports multiple multi-modal proxies (images, text) as well as missing data. We empirically demonstrate that our approach outperforms existing methods based on propensity scores and corrects for confounding using unstructured inputs on tasks in genomics and healthcare. Our methods can potentially support the use of large amounts of data that were previously not used in causal inference

Shachi Deshpande, Volodymyr Kuleshov · microsoft-research

Propensity scores are commonly used to estimate treatment effects from observational data. We argue that the probabilistic output of a learned propensity score model should be calibrated -- i.e., a predictive treatment probability of 90% should correspond to 90% of individuals being assigned the treatment group -- and we propose simple recalibration techniques to ensure this property. We prove that calibration is a necessary condition for unbiased treatment effect estimation when using popular inverse propensity weighted and doubly robust estimators. We derive error bounds on causal effect estimates that directly relate to the quality of uncertainties provided by the probabilistic propensity score model and show that calibration strictly improves this error bound while also avoiding extreme propensity weights. We demonstrate improved causal effect estimation with calibrated propensity scores in several tasks including high-dimensional image covariates and genome-wide association studies (GWASs). Calibrated propensity scores improve the speed of GWAS analysis by more than two-fold by enabling the use of simpler models that are faster to train.

Shachi Deshpande, Charles Marx, Volodymyr Kuleshov · microsoft-research

We are interested in probabilistic prediction in online settings in which data does not follow a probability distribution. Our work seeks to achieve two goals: (1) producing valid probabilities that accurately reflect model confidence; and (2) ensuring that traditional notions of performance (e.g., high accuracy) still hold. We introduce online algorithms guaranteed to achieve these goals on arbitrary streams of data points, including data chosen by an adversary. Specifically, our algorithms produce forecasts that are (1) calibrated -- i.e., an 80% confidence interval contains the true outcome 80% of the time -- and (2) have low regret relative to a user-specified baseline model. We implement a post-hoc recalibration strategy that provably achieves these goals in regression; previous algorithms applied to classification or achieved (1) but not (2). In the context of Bayesian optimization, an online model-based decision-making task in which the data distribution shifts over time, our method yields accelerated convergence to improved optima.

Charles Marx, Volodymyr Kuleshov, Stefano Ermon

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.

Subham Sekhar Sahoo, Anselm Paulus, Marin Vlastelica et al.

Embedding discrete solvers as differentiable layers has given modern deep learning architectures combinatorial expressivity and discrete reasoning capabilities. The derivative of these solvers is zero or undefined, therefore a meaningful replacement is crucial for effective gradient-based learning. Prior works rely on smoothing the solver with input perturbations, relaxing the solver to continuous problems, or interpolating the loss landscape with techniques that typically require additional solver calls, introduce extra hyper-parameters, or compromise performance. We propose a principled approach to exploit the geometry of the discrete solution space to treat the solver as a negative identity on the backward pass and further provide a theoretical justification. Our experiments demonstrate that such a straightforward hyper-parameter-free approach is able to compete with previous more complex methods on numerous experiments such as backpropagation through discrete samplers, deep graph matching, and image retrieval. Furthermore, we substitute the previously proposed problem-specific and label-dependent margin with a generic regularization procedure that prevents cost collapse and increases robustness.

Jacqueline R. M. A. Maasch, Hao Zhang, Qian Yang et al.

The cost of manual data labeling can be a significant obstacle in supervised learning. Data programming (DP) offers a weakly supervised solution for training dataset creation, wherein the outputs of user-defined programmatic labeling functions (LFs) are reconciled through unsupervised learning. However, DP can fail to outperform an unweighted majority vote in some scenarios, including low-data contexts. This work introduces a Bayesian extension of classical DP that mitigates failures of unsupervised learning by augmenting the DP objective with regularization terms. Regularized learning is achieved through maximum a posteriori estimation with informative priors. Majority vote is proposed as a proxy signal for automated prior parameter selection. Results suggest that regularized DP improves performance relative to maximum likelihood and majority voting, confers greater interpretability, and bolsters performance in low-data regimes.

Phillip Si, Zeyi Chen, Subham Sekhar Sahoo et al.

Training normalizing flow generative models can be challenging due to the need to calculate computationally expensive determinants of Jacobians. This paper studies the likelihood-free training of flows and proposes the energy objective, an alternative sample-based loss based on proper scoring rules. The energy objective is determinant-free and supports flexible model architectures that are not easily compatible with maximum likelihood training, including semi-autoregressive energy flows, a novel model family that interpolates between fully autoregressive and non-autoregressive models. Energy flows feature competitive sample quality, posterior inference, and generation speed relative to likelihood-based flows; this performance is decorrelated from the quality of log-likelihood estimates, which are generally very poor. Our findings question the use of maximum likelihood as an objective or a metric, and contribute to a scientific study of its role in generative modeling.

Albert Tseng, Jerry Chee, Qingyao Sun et al.

Post-training quantization (PTQ) reduces the memory footprint of LLMs by quantizing their weights to low-precision. In this work, we introduce QuIP#, a weight-only PTQ method that achieves state-of-the-art results in extreme compression regimes ($\le$ 4 bits per weight) using three novel techniques. First, QuIP# improves QuIP's (Chee et al., 2023) incoherence processing by using the randomized Hadamard transform, which is faster and has better theoretical properties. Second, QuIP# uses vector quantization to take advantage of the ball-shaped sub-Gaussian distribution that incoherent weights possess: specifically, we introduce a set of hardware-efficient codebooks based on the highly symmetric $E_8$ lattice, which achieves the optimal 8-dimension unit ball packing. Third, QuIP# uses fine-tuning to improve fidelity to the original model. Our experiments show that QuIP# outperforms existing PTQ methods, enables new behaviors in PTQ scaling, and supports fast inference. Our code can be found at https://github.com/Cornell-RelaxML/quip-sharp.

Jacqueline Maasch, Weishen Pan, Shantanu Gupta et al.

Causal discovery is crucial for causal inference in observational studies, as it can enable the identification of valid adjustment sets (VAS) for unbiased effect estimation. However, global causal discovery is notoriously hard in the nonparametric setting, with exponential time and sample complexity in the worst case. To address this, we propose local discovery by partitioning (LDP): a local causal discovery method that is tailored for downstream inference tasks without requiring parametric and pretreatment assumptions. LDP is a constraint-based procedure that returns a VAS for an exposure-outcome pair under latent confounding, given sufficient conditions. The total number of independence tests performed is worst-case quadratic with respect to the cardinality of the variable set. Asymptotic theoretical guarantees are numerically validated on synthetic graphs. Adjustment sets from LDP yield less biased and more precise average treatment effect estimates than baseline discovery algorithms, with LDP outperforming on confounder recall, runtime, and test count for VAS discovery. Notably, LDP ran at least 1300x faster than baselines on a benchmark.

Junjie Yin, Jiahao Dong, Yingheng Wang et al.

We propose a memory-efficient finetuning algorithm for large language models (LLMs) that supports finetuning LLMs with 65B parameters in 2/3/4-bit precision on as little as one 24GB GPU. Our method, modular low-rank adaptation (ModuLoRA), integrates any user-specified weight quantizer with finetuning via low-rank adapters (LoRAs). Our approach relies on a simple quantization-agnostic backward pass that adaptively materializes low-precision LLM weights from a custom black-box quantization module. This approach enables finetuning 2-bit and 3-bit LLMs for the first time -- leveraging state-of-the-art 2-bit QuIP\# quantization and 3-bit OPTQ quantization -- outperforming finetuning that relies on less sophisticated 4-bit and 8-bit methods. In our experiments, \lplora~attains competitive performance on text classification, natural language inference, and instruction following tasks using significantly less memory than existing approaches, and we also surpass the state-of-the-art ROUGE score on a popular summarization task. We release \lplora~together with a series of low-precision models as part of \llmtune, a user-friendly library for quantizing, running, and finetuning LLMs on consumer GPUs.

Gilad Turok, Chris De Sa, Volodymyr Kuleshov

Masked diffusion models (MDMs) generate text by iteratively selecting positions to unmask and then predicting tokens at those positions. Yet MDMs lack proper perplexity evaluation: the ELBO is a loose bound on likelihood under the training distribution, not the test-time distribution, while generative perplexity requires a biased external model and ignores diversity. To address this, we introduce the \textsc{DUEL} framework, which formalizes \emph{deterministic} position selection, unifying leading MDM sampling strategies. We prove \textbf{\textsc{DUEL} admits \emph{exact} likelihood computation} via a simple algorithm, evaluated under the same position selection used at test time. This \textbf{gives MDMs proper perplexity for the first time} -- the natural analogue of autoregressive perplexity. With proper perplexity in hand, we revisit key questions about MDMs. \textbf{MDMs are substantially better than previously thought}: the MDM-autoregressive perplexity gap shrinks by up to 32\% on in-domain data and 82\% on zero-shot benchmarks. \textsc{DUEL} enables the first principled comparison of fast, parallel samplers across compute budgets -- an analysis impossible with the ELBO and unreliable with generative perplexity -- identifying probability margin \citep{kim2025train} as a strong default. Finally, oracle search over position orderings reveals MDMs can far surpass autoregressive models -- achieving 36.47 vs.\ 52.11 perplexity on AG News -- demonstrating the ceiling of MDM performance has not yet been reached.

Yair Schiff, Omer Belhasin, Roy Uziel et al.

Masked diffusion models (MDMs) have emerged as a promising alternative to autoregressive models, enabling parallel token generation while achieving competitive performance. Despite these advantages, MDMs face a fundamental limitation: once tokens are unmasked, they remain fixed, leading to error accumulation and ultimately degrading sample quality. We address this by proposing a framework that trains a model to perform both unmasking and correction. By reusing outputs from the MDM denoising network as inputs for corrector training, we train a model to recover from potential mistakes. During generation we apply additional corrective refinement steps between unmasking ones in order to change decoded tokens and improve outputs. We name our training and sampling method Progressive Self-Correction (ProSeCo) for its unique ability to iteratively refine an entire sequence, including already generated tokens. We conduct extensive experimental validation across multiple conditional and unconditional tasks, demonstrating that ProSeCo yields better quality-efficiency trade-offs (up to ~2-3x faster sampling) and enables inference-time compute scaling to further increase sample quality beyond standard MDMs (up to ~1.3x improvement on benchmarks).

Yair Schiff, Chia-Hsiang Kao, Aaron Gokaslan et al.

Large-scale sequence modeling has sparked rapid advances that now extend into biology and genomics. However, modeling genomic sequences introduces challenges such as the need to model long-range token interactions, the effects of upstream and downstream regions of the genome, and the reverse complementarity (RC) of DNA. Here, we propose an architecture motivated by these challenges that builds off the long-range Mamba block, and extends it to a BiMamba component that supports bi-directionality, and to a MambaDNA block that additionally supports RC equivariance. We use MambaDNA as the basis of Caduceus, the first family of RC equivariant bi-directional long-range DNA language models, and we introduce pre-training and fine-tuning strategies that yield Caduceus DNA foundation models. Caduceus outperforms previous long-range models on downstream benchmarks; on a challenging long-range variant effect prediction task, Caduceus exceeds the performance of 10x larger models that do not leverage bi-directionality or equivariance.

Marianne Arriola, Aaron Gokaslan, Justin T. Chiu et al.

Diffusion language models offer unique benefits over autoregressive models due to their potential for parallelized generation and controllability, yet they lag in likelihood modeling and are limited to fixed-length generation. In this work, we introduce a class of block diffusion language models that interpolate between discrete denoising diffusion and autoregressive models. Block diffusion overcomes key limitations of both approaches by supporting flexible-length generation and improving inference efficiency with KV caching and parallel token sampling. We propose a recipe for building effective block diffusion models that includes an efficient training algorithm, estimators of gradient variance, and data-driven noise schedules to minimize the variance. Block diffusion sets a new state-of-the-art performance among diffusion models on language modeling benchmarks and enables generation of arbitrary-length sequences. We provide the code, along with the model weights and blog post on the project page: https://m-arriola.com/bd3lms

Inception Labs, Samar Khanna, Siddhant Kharbanda et al. · deepmind, microsoft-research

We present Mercury, a new generation of commercial-scale large language models (LLMs) based on diffusion. These models are parameterized via the Transformer architecture and trained to predict multiple tokens in parallel. In this report, we detail Mercury Coder, our first set of diffusion LLMs designed for coding applications. Currently, Mercury Coder comes in two sizes: Mini and Small. These models set a new state-of-the-art on the speed-quality frontier. Based on independent evaluations conducted by Artificial Analysis, Mercury Coder Mini and Mercury Coder Small achieve state-of-the-art throughputs of 1109 tokens/sec and 737 tokens/sec, respectively, on NVIDIA H100 GPUs and outperform speed-optimized frontier models by up to 10x on average while maintaining comparable quality. We discuss additional results on a variety of code benchmarks spanning multiple languages and use-cases as well as real-world validation by developers on Copilot Arena, where the model currently ranks second on quality and is the fastest model overall. We also release a public API at https://platform.inceptionlabs.ai/ and free playground at https://chat.inceptionlabs.ai

Guanghan Wang, Yair Schiff, Subham Sekhar Sahoo et al.

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://remdm.github.io

Yair Schiff, Subham Sekhar Sahoo, Hao Phung et al.

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.

Yiwen Huang, Aaron Gokaslan, Volodymyr Kuleshov et al.

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.

Subham Sekhar Sahoo, Aaron Gokaslan, Chris De Sa et al.

Diffusion models have gained traction as powerful algorithms for synthesizing high-quality images. Central to these algorithms is the diffusion process, a set of equations which maps data to noise in a way that can significantly affect performance. In this paper, we explore whether the diffusion process can be learned from data. Our work is grounded in Bayesian inference and seeks to improve log-likelihood estimation by casting the learned diffusion process as an approximate variational posterior that yields a tighter lower bound (ELBO) on the likelihood. A widely held assumption is that the ELBO is invariant to the noise process: our work dispels this assumption and proposes multivariate learned adaptive noise (MULAN), a learned diffusion process that applies noise at different rates across an image. Specifically, our method relies on a multivariate noise schedule that is a function of the data to ensure that the ELBO is no longer invariant to the choice of the noise schedule as in previous works. Empirically, MULAN sets a new state-of-the-art in density estimation on CIFAR-10 and ImageNet and reduces the number of training steps by 50%. We provide the code, along with a blog post and video tutorial on the project page: https://s-sahoo.com/MuLAN

Subham Sekhar Sahoo, Justin Deschenaux, Aaron Gokaslan et al.

Uniform-state discrete diffusion models hold the promise of fast text generation due to their inherent ability to self-correct. However, they are typically outperformed by autoregressive models and masked diffusion models. In this work, we narrow this performance gap by leveraging a key insight: Uniform-state diffusion processes naturally emerge from an underlying Gaussian diffusion. Our method, Duo, transfers powerful techniques from Gaussian diffusion to improve both training and sampling. First, we introduce a curriculum learning strategy guided by the Gaussian process, doubling training speed by reducing variance. Models trained with curriculum learning surpass autoregressive models in zero-shot perplexity on 3 of 7 benchmarks. Second, we present Discrete Consistency Distillation, which adapts consistency distillation from the continuous to the discrete setting. This algorithm unlocks few-step generation in diffusion language models by accelerating sampling by two orders of magnitude. We provide the code and model checkpoints on the project page: http://s-sahoo.github.io/duo

Wasu Top Piriyakulkij, Volodymyr Kuleshov, Kevin Ellis

Actively inferring user preferences, for example by asking good questions, is important for any human-facing decision-making system. Active inference allows such systems to adapt and personalize themselves to nuanced individual preferences. To enable this ability for instruction-tuned large language models (LLMs), one may prompt them to ask users questions to infer their preferences, transforming the language models into more robust, interactive systems. However, out of the box, these models are not efficient at extracting preferences: the questions they generate are not informative, requiring a high number of user interactions and impeding the usability of the downstream system. In this work, we introduce an inference-time algorithm that helps LLMs quickly infer preferences by using more informative questions. Our algorithm uses a probabilistic model whose conditional distributions are defined by prompting an LLM, and returns questions that optimize expected entropy and expected model change. Results in a simplified interactive web shopping setting with real product items show that an LLM equipped with our entropy reduction algorithm outperforms baselines with the same underlying LLM on task performance while using fewer user interactions.

Yair Schiff, Zhong Yi Wan, Jeffrey B. Parker et al.

Learning dynamics from dissipative chaotic systems is notoriously difficult due to their inherent instability, as formalized by their positive Lyapunov exponents, which exponentially amplify errors in the learned dynamics. However, many of these systems exhibit ergodicity and an attractor: a compact and highly complex manifold, to which trajectories converge in finite-time, that supports an invariant measure, i.e., a probability distribution that is invariant under the action of the dynamics, which dictates the long-term statistical behavior of the system. In this work, we leverage this structure to propose a new framework that targets learning the invariant measure as well as the dynamics, in contrast with typical methods that only target the misfit between trajectories, which often leads to divergence as the trajectories' length increases. We use our framework to propose a tractable and sample efficient objective that can be used with any existing learning objectives. Our Dynamics Stable Learning by Invariant Measure (DySLIM) objective enables model training that achieves better point-wise tracking and long-term statistical accuracy relative to other learning objectives. By targeting the distribution with a scalable regularization term, we hope that this approach can be extended to more complex systems exhibiting slowly-variant distributions, such as weather and climate models.

Guanghan Wang, Yair Schiff, Gilad Turok et al.

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 properties of masking to accurately estimate the likelihoods of sampling trajectories. Our estimators trade off computation for approximation accuracy in an analytically tractable manner, and are particularly effective for DLMs that support any-order likelihood estimation. We characterize and study this property in popular DLMs and show that it is key for efficient diffusion-based reasoning. Empirically, d2 significantly improves over previous diffusion reasoning frameworks using only RL (without relying on supervised fine-tuning), 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).

Wasu Top Piriyakulkij, Yingheng Wang, Volodymyr Kuleshov

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 the Thousand Genomes dataset.

Marianne Arriola, Yair Schiff, Hao Phung et al.

Discrete diffusion models enable parallel token sampling for faster inference than autoregressive approaches. However, prior diffusion models use a decoder-only architecture, which requires sampling algorithms that invoke the full network at every denoising step and incur high computational cost. Our key insight is that discrete diffusion models perform two types of computation: 1) representing clean tokens and 2) denoising corrupted tokens, which enables us to use separate modules for each task. We propose an encoder-decoder architecture to accelerate discrete diffusion inference, which relies on an encoder to represent clean tokens and a lightweight decoder to iteratively refine a noised sequence. We also show that this architecture enables faster training of block diffusion models, which partition sequences into blocks for better quality and are commonly used in diffusion language model inference. We introduce a framework for Efficient Encoder-Decoder Diffusion (E2D2), consisting of an architecture with specialized training and sampling algorithms, and we show that E2D2 achieves superior trade-offs between generation quality and inference throughput on summarization, translation, and mathematical reasoning tasks. We provide the code, model weights, and blog post on the project page: https://m-arriola.com/e2d2

Jacqueline Maasch, Willie Neiswanger, Stefano Ermon et al.

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.

Subham Sekhar Sahoo, Marianne Arriola, Yair Schiff et al.

While diffusion models excel at generating high-quality images, prior work reports a significant performance gap between diffusion and autoregressive (AR) methods in language modeling. In this work, we show that simple masked discrete diffusion is more performant than previously thought. We apply an effective training recipe that improves the performance of masked diffusion models and derive a simplified, Rao-Blackwellized objective that results in additional improvements. Our objective has a simple form -- it is a mixture of classical masked language modeling losses -- and can be used to train encoder-only language models that admit efficient samplers, including ones that can generate arbitrary lengths of text semi-autoregressively like a traditional language model. On language modeling benchmarks, a range of masked diffusion models trained with modern engineering practices achieves a new state-of-the-art among diffusion models, and approaches AR perplexity. We provide the code, along with a blog post and video tutorial on the project page: https://s-sahoo.com/mdlm

Volodymyr Kuleshov, Shachi Deshpande

Accurate probabilistic predictions can be characterized by two properties -- calibration and sharpness. However, standard maximum likelihood training yields models that are poorly calibrated and thus inaccurate -- a 90% confidence interval typically does not contain the true outcome 90% of the time. This paper argues that calibration is important in practice and is easy to maintain by performing low-dimensional density estimation. We introduce a simple training procedure based on recalibration that yields calibrated models without sacrificing overall performance; unlike previous approaches, ours ensures the most general property of distribution calibration and applies to any model, including neural networks. We formally prove the correctness of our procedure assuming that we can estimate densities in low dimensions and we establish uniform convergence bounds. Our results yield empirical performance improvements on linear and deep Bayesian models and suggest that calibration should be increasingly leveraged across machine learning. We release a library that implements our methods along with a blog post here: https://shachideshpande.github.io/blog-distribution-calibration/.

Volodymyr Kuleshov, Evgenii Nikishin, Shantanu Thakoor et al.

Modern classification algorithms are susceptible to adversarial examples--perturbations to inputs that cause the algorithm to produce undesirable behavior. In this work, we seek to understand and extend adversarial examples across domains in which inputs are discrete, particularly across new domains, such as computational biology. As a step towards this goal, we formalize a notion of synonymous adversarial examples that applies in any discrete setting and describe a simple domain-agnostic algorithm to construct such examples. We apply this algorithm across multiple domains--including sentiment analysis and DNA sequence classification--and find that it consistently uncovers adversarial examples. We seek to understand their prevalence theoretically and we attribute their existence to spurious token correlations, a statistical phenomenon that is specific to discrete spaces. Our work is a step towards a domain-agnostic treatment of discrete adversarial examples analogous to that of continuous inputs.

Phillip Si, Allan Bishop, Volodymyr Kuleshov

Numerous applications of machine learning involve representing probability distributions over high-dimensional data. We propose autoregressive quantile flows, a flexible class of normalizing flow models trained using a novel objective based on proper scoring rules. Our objective does not require calculating computationally expensive determinants of Jacobians during training and supports new types of neural architectures, such as neural autoregressive flows from which it is easy to sample. We leverage these models in quantile flow regression, an approach that parameterizes predictive conditional distributions with flows, resulting in improved probabilistic predictions on tasks such as time series forecasting and object detection. Our novel objective functions and neural flow parameterizations also yield improvements on popular generation and density estimation tasks, and represent a step beyond maximum likelihood learning of flows.

Shachi Deshpande, Charles Marx, Volodymyr Kuleshov

Accurate uncertainty estimates are important in sequential model-based decision-making tasks such as Bayesian optimization. However, these estimates can be imperfect if the data violates assumptions made by the model (e.g., Gaussianity). This paper studies which uncertainties are needed in model-based decision-making and in Bayesian optimization, and argues that uncertainties can benefit from calibration -- i.e., an 80% predictive interval should contain the true outcome 80% of the time. Maintaining calibration, however, can be challenging when the data is non-stationary and depends on our actions. We propose using simple algorithms based on online learning to provably maintain calibration on non-i.i.d. data, and we show how to integrate these algorithms in Bayesian optimization with minimal overhead. Empirically, we find that calibrated Bayesian optimization converges to better optima in fewer steps, and we demonstrate improved performance on standard benchmark functions and hyperparameter optimization tasks.

Bojian Hou, Hao Zhang, Gur Ladizhinsky et al.

Abstruse learning algorithms and complex datasets increasingly characterize modern clinical decision support systems (CDSS). As a result, clinicians cannot easily or rapidly scrutinize the CDSS recommendation when facing a difficult diagnosis or treatment decision in practice. Over-trust or under-trust are frequent. Prior research has explored supporting such assessments by explaining DST data inputs and algorithmic mechanisms. This paper explores a different approach: Providing precisely relevant, scientific evidence from biomedical literature. We present a proof-of-concept system, Clinical Evidence Engine, to demonstrate the technical and design feasibility of this approach across three domains (cardiovascular diseases, autism, cancer). Leveraging Clinical BioBERT, the system can effectively identify clinical trial reports based on lengthy clinical questions (e.g., "risks of catheter infection among adult patients in intensive care unit who require arterial catheters, if treated with povidone iodine-alcohol"). This capability enables the system to identify clinical trials relevant to diagnostic/treatment hypotheses -- a clinician's or a CDSS's. Further, Clinical Evidence Engine can identify key parts of a clinical trial abstract, including patient population (e.g., adult patients in intensive care unit who require arterial catheters), intervention (povidone iodine-alcohol), and outcome (risks of catheter infection). This capability opens up the possibility of enabling clinicians to 1) rapidly determine the match between a clinical trial and a clinical question, and 2) understand the result and contexts of the trial without extensive reading. We demonstrate this potential by illustrating two example use scenarios of the system. We discuss the idea of designing DST explanations not as specific to a DST or an algorithm, but as a domain-agnostic decision support infrastructure.

Sawyer Birnbaum, Volodymyr Kuleshov, Zayd Enam et al.

Learning representations that accurately capture long-range dependencies in sequential inputs -- including text, audio, and genomic data -- is a key problem in deep learning. Feed-forward convolutional models capture only feature interactions within finite receptive fields while recurrent architectures can be slow and difficult to train due to vanishing gradients. Here, we propose Temporal Feature-Wise Linear Modulation (TFiLM) -- a novel architectural component inspired by adaptive batch normalization and its extensions -- that uses a recurrent neural network to alter the activations of a convolutional model. This approach expands the receptive field of convolutional sequence models with minimal computational overhead. Empirically, we find that TFiLM significantly improves the learning speed and accuracy of feed-forward neural networks on a range of generative and discriminative learning tasks, including text classification and audio super-resolution

Ali Malik, Volodymyr Kuleshov, Jiaming Song et al.

Estimates of predictive uncertainty are important for accurate model-based planning and reinforcement learning. However, predictive uncertainties---especially ones derived from modern deep learning systems---can be inaccurate and impose a bottleneck on performance. This paper explores which uncertainties are needed for model-based reinforcement learning and argues that good uncertainties must be calibrated, i.e. their probabilities should match empirical frequencies of predicted events. We describe a simple way to augment any model-based reinforcement learning agent with a calibrated model and show that doing so consistently improves planning, sample complexity, and exploration. On the \textsc{HalfCheetah} MuJoCo task, our system achieves state-of-the-art performance using 50\% fewer samples than the current leading approach. Our findings suggest that calibration can improve the performance of model-based reinforcement learning with minimal computational and implementation overhead.

Volodymyr Kuleshov, Nathan Fenner, Stefano Ermon

Methods for reasoning under uncertainty are a key building block of accurate and reliable machine learning systems. Bayesian methods provide a general framework to quantify uncertainty. However, because of model misspecification and the use of approximate inference, Bayesian uncertainty estimates are often inaccurate -- for example, a 90% credible interval may not contain the true outcome 90% of the time. Here, we propose a simple procedure for calibrating any regression algorithm; when applied to Bayesian and probabilistic models, it is guaranteed to produce calibrated uncertainty estimates given enough data. Our procedure is inspired by Platt scaling and extends previous work on classification. We evaluate this approach on Bayesian linear regression, feedforward, and recurrent neural networks, and find that it consistently outputs well-calibrated credible intervals while improving performance on time series forecasting and model-based reinforcement learning tasks.

Hongyu Ren, Russell Stewart, Jiaming Song et al.

Constraint-based learning reduces the burden of collecting labels by having users specify general properties of structured outputs, such as constraints imposed by physical laws. We propose a novel framework for simultaneously learning these constraints and using them for supervision, bypassing the difficulty of using domain expertise to manually specify constraints. Learning requires a black-box simulator of structured outputs, which generates valid labels, but need not model their corresponding inputs or the input-label relationship. At training time, we constrain the model to produce outputs that cannot be distinguished from simulated labels by adversarial training. Providing our framework with a small number of labeled inputs gives rise to a new semi-supervised structured prediction model; we evaluate this model on multiple tasks --- tracking, pose estimation and time series prediction --- and find that it achieves high accuracy with only a small number of labeled inputs. In some cases, no labels are required at all.

Volodymyr Kuleshov, Stefano Ermon

Many problems in machine learning are naturally expressed in the language of undirected graphical models. Here, we propose black-box learning and inference algorithms for undirected models that optimize a variational approximation to the log-likelihood of the model. Central to our approach is an upper bound on the log-partition function parametrized by a function q that we express as a flexible neural network. Our bound makes it possible to track the partition function during learning, to speed-up sampling, and to train a broad class of hybrid directed/undirected models via a unified variational inference framework. We empirically demonstrate the effectiveness of our method on several popular generative modeling datasets.

Volodymyr Kuleshov, S. Zayd Enam, Stefano Ermon

We introduce a new audio processing technique that increases the sampling rate of signals such as speech or music using deep convolutional neural networks. Our model is trained on pairs of low and high-quality audio examples; at test-time, it predicts missing samples within a low-resolution signal in an interpolation process similar to image super-resolution. Our method is simple and does not involve specialized audio processing techniques; in our experiments, it outperforms baselines on standard speech and music benchmarks at upscaling ratios of 2x, 4x, and 6x. The method has practical applications in telephony, compression, and text-to-speech generation; it demonstrates the effectiveness of feed-forward convolutional architectures on an audio generation task.

Volodymyr Kuleshov, Stefano Ermon

Assessing uncertainty is an important step towards ensuring the safety and reliability of machine learning systems. Existing uncertainty estimation techniques may fail when their modeling assumptions are not met, e.g. when the data distribution differs from the one seen at training time. Here, we propose techniques that assess a classification algorithm's uncertainty via calibrated probabilities (i.e. probabilities that match empirical outcome frequencies in the long run) and which are guaranteed to be reliable (i.e. accurate and calibrated) on out-of-distribution input, including input generated by an adversary. This represents an extension of classical online learning that handles uncertainty in addition to guaranteeing accuracy under adversarial assumptions. We establish formal guarantees for our methods, and we validate them on two real-world problems: question answering and medical diagnosis from genomic data.

Volodymyr Kuleshov, Arun Tejasvi Chaganty, Percy Liang

Tensor factorization arises in many machine learning applications, such knowledge base modeling and parameter estimation in latent variable models. However, numerical methods for tensor factorization have not reached the level of maturity of matrix factorization methods. In this paper, we propose a new method for CP tensor factorization that uses random projections to reduce the problem to simultaneous matrix diagonalization. Our method is conceptually simple and also applies to non-orthogonal and asymmetric tensors of arbitrary order. We prove that a small number random projections essentially preserves the spectral information in the tensor, allowing us to remove the dependence on the eigengap that plagued earlier tensor-to-matrix reductions. Experimentally, our method outperforms existing tensor factorization methods on both simulated data and two real datasets.

Volodymyr Kuleshov, Doina Precup

Although many algorithms for the multi-armed bandit problem are well-understood theoretically, empirical confirmation of their effectiveness is generally scarce. This paper presents a thorough empirical study of the most popular multi-armed bandit algorithms. Three important observations can be made from our results. Firstly, simple heuristics such as epsilon-greedy and Boltzmann exploration outperform theoretically sound algorithms on most settings by a significant margin. Secondly, the performance of most algorithms varies dramatically with the parameters of the bandit problem. Our study identifies for each algorithm the settings where it performs well, and the settings where it performs poorly. Thirdly, the algorithms' performance relative each to other is affected only by the number of bandit arms and the variance of the rewards. This finding may guide the design of subsequent empirical evaluations. In the second part of the paper, we turn our attention to an important area of application of bandit algorithms: clinical trials. Although the design of clinical trials has been one of the principal practical problems motivating research on multi-armed bandits, bandit algorithms have never been evaluated as potential treatment allocation strategies. Using data from a real study, we simulate the outcome that a 2001-2002 clinical trial would have had if bandit algorithms had been used to allocate patients to treatments. We find that an adaptive trial would have successfully treated at least 50% more patients, while significantly reducing the number of adverse effects and increasing patient retention. At the end of the trial, the best treatment could have still been identified with a high level of statistical confidence. Our findings demonstrate that bandit algorithms are attractive alternatives to current adaptive treatment allocation strategies.