Zachary Izzo is a Researcher in the Machine Learning Department at NEC Laboratories America in Princeton, NJ. He received his PhD in Mathematics from Stanford University and his BS in Mathematics from the Massachusetts Institute of Technology. His research develops mathematically rigorous methods to understand when and why machine learning models work, with a recent focus on the foundations and interpretability of large language models.
Recent work includes a provably convergent algorithm for identifying patient subgroups with differential treatment response, and theoretical results on length generalization in transformers. By advancing methods that enhance interpretability and adaptability, his research directly contributes to NEC’s mission to build intelligent, resilient AI technologies that can operate in complex, data-rich real-world settings. His contributions help extend the frontier of AI toward systems that are not only accurate but also dependable and actionable in practice.
NEC Laboratories America researchers are heading to Seoul this July for ICML 2026, the Forty-Third International Conference on Machine Learning. One of the most prestigious gatherings in the field, ICML draws academic and industry researchers from around the world to share work spanning machine learning, artificial intelligence, data science, and their many applications.
Diffusion models (DMs) have proven to be effective in modeling high-dimensional distributions, leading to their widespread adoption for representing complex priors in Bayesian inverse problems (BIPs). However, current DM-based posterior sampling methods proposed for solving common BIPs rely on heuristic approximations to the generative process. To exploit the generative capability of DMs and avoid the usage of such approximations, we propose an ensemble-based algorithm that performs posterior sampling without the use of heuristic approximations. Our algorithm is motivated by existing work that combines DM-based methods with the sequential Monte Carlo (SMC) method. By examining how the prior evolves through the diffusion process encoded by the pre-trained score function, we derive a modified partial differential equation (PDE) governing the evolution of the corresponding posterior distribution. This PDE includes a modified diffusion term and a reweighting term, which can be simulated via stochastic weighted particle methods. Theoretically, we prove that the error between the true posterior and the empirical distribution of the generated samples can be bounded in terms of the training error of the pre-trained score function and the number of particles in the ensemble. Empirically, we validate our algorithm on several inverse problems in imaging to show that our method gives more accurate reconstructions compared to existing DM-based methods.
We study the problem of length generalization (LG) in transformers: the ability of a model trained on shorter sequences to maintain performance when evaluated on much longer, previously unseen inputs. Prior work by Huang et al. (2024) established that transformers eventually achieve length generalization once the training sequence length exceeds some finite threshold, but left open the question of how large it must be. In this work, we provide the first quantitative bounds on the required training length for length generalization to occur. Motivated by previous empirical and theoretical work, we analyze LG in several distinct problem settings: error control vs. average error control over an input distribution, infinite-precision softmax attention vs. finite-precision attention (which reduces to an argmax) in the transformer, as well as for one- or two-layer transformers. In all scenarios, we prove that LG occurs when the internal behavior of the transformer on longer sequences can be “simulated” by its behavior on shorter sequences seen during training. Our bounds give qualitative estimates for the required length of training data required for a transformer to generalize, and we verify these insights empirically. These results sharpen our theoretical understanding of the mechanisms underlying extrapolation in transformers, and formalize the intuition that richer training data is required for generalization on more complex tasks.
How many mistakes do published AI papers contain? Peer-reviewed publications form the foundation upon which new research and knowledge are built. Errors that persist in the literature can propagate unnoticed, creating confusion in follow-up studies and complicating reproducibility. The accelerating pace of research and the increasing demands on the peer-review system make such mistakes harder to detect and avoid. To address this, we developed a Paper Correctness Checker based on GPT-5 to systematically identify mistakes in papers previously published at top AI conferences and journals. Our analysis focuses on objective mistakes-e.g., errors in formulas, derivations, calculations, figures, and tables-that have a clearly verifiable ground truth. We intentionally exclude subjective considerations such as novelty, importance, or writing quality. We find that published papers contain a non-negligible number of objective mistakes and that the average number of mistakes per paper has increased over time-from 3.8 in NeurIPS 2021 to 5.9 in NeurIPS 2025 (55.3% increase); from 4.1 in ICLR 2018 to 5.2 in ICLR 2025; and from 5.0 in TMLR 2022/23 to 5.5 in TMLR 2025. Human experts reviewed 316 potential mistakes identified by the AI Checker and confirmed that 263 were actual mistakes, corresponding to a precision of 83.2%. While most identified issues are relatively minor, correcting them would reduce confusion in the literature and strengthen reproducibility. The AI Checker also surfaced potentially more substantive mistakes that could affect the interpretation of results. Moreover, we show that the AI Checker can propose correct fixes for 75.8% of the identified mistakes. Overall, this study highlights the potential of frontier LLMs to detect and correct objective mistakes in published papers, helping to establish a firmer foundation of knowledge.
We study the problem of length generalization (LG) in transformers: the ability of a model trained on shorter sequences to maintain performance when evaluated on much longer, previously unseen inputs. Prior work by Huang et al. (2025) established that transformers eventually achieve length generalization once the training sequence length exceeds some finite threshold, but left open the question of how large it must be. In this work, we provide the first quantitative bounds on the required training length for length generalization to occur. Motivated by previous empirical and theoretical work, we analyze LG in several distinct problem settings: error control vs. average error control over an input distribution, infinite-precision softmax attention vs. finite-precision attention (which reduces to an argmax) in the transformer, and one- vs. two-layer transformers. In all scenarios, we prove that LG occurs when the internal behavior of the transformer on longer sequences can be “simulated” by its behavior on shorter sequences seen during training. Our bounds give qualitative estimates for the length of training data required for a transformer to generalize, and we verify these insights empirically. These results sharpen our theoretical understanding of the mechanisms underlying extrapolation in transformers, and formalize the intuition that richer training data is required for generalization on more complex tasks.
Adapting large Video-Language Models (VLMs) for action detection using only a few examples poses challenges like overfitting and the granularity mismatch between scene-level pre-training and required person-centric understanding. We propose an efficient adaptation strategy combining parameter-efficient tuning (LoRA) with a novel learnable internal feature augmentation. Applied within the frozen VLM backbone using FiLM, these augmentations generate diverse feature variations directly relevant to the task. Additionally, we introduce a group-weighted loss function that dynamically modulates the training contribution of each augmented sample based on its prediction divergence relative to the group average. This promotes robust learning by prioritizing informative yet reasonable augmentations. We demonstrate our method’s effectiveness on complex multi-label, multi-person action detection datasets (AVA, MOMA), achieving strong mAP performance and showcasing significant data efficiency for adapting VLMs from limited examples.
We provide quantitative bounds on the length of sequences required to be observed during training for a transformer to length generalize, e.g., to continue to perform well on sequences unseen during training. Our results improve on Huang et al. [8], who show that there is a finite training length beyond which length generalization is guaranteed, but for which they do not provide quantitative bounds.
Diffusion models (DMs) have proven to be effective in modeling high-dimensional distributions, leading to their widespread adoption for representing complex priors in Bayesian inverse problems (BIPs). However, current DM-based posterior sampling methods proposed for solving common BIPs rely on heuristic approximations to the generative process. To exploit the generative capability of DMs and avoid the usage of such approximations, we propose an ensemble-based algorithm that performs posterior sampling without the use of heuristic approximations. Our algorithm is motivated by existing works that combine DM-based methods with the sequential Monte Carlo (SMC) method. By examining how the prior evolves through the diffusion process encoded by the pre-trained score function, we derive a modified partial differential equation (PDE) governing the evolution of the corresponding posterior distribution. This PDE includes a modified diffusion term and a reweighting term, which can be simulated via stochastic weighted particle methods. Theoretically, we prove that the error between the true posterior distribution canbe bounded in terms of the training error of the pre-trained score function and the ]number of particles in the ensemble. Empirically, we validate our algorithm on several inverse problems in imaging to show that our method gives more accurate reconstructions compared to existing DM-based methods.
Unarguably deep learning models capable of generalizing to unseen domain data while leveraging a few labels are of great practical significance due to low developmental costs. In search of this endeavor we study the challenging problem of semi-supervised domain generalization (SSDG) where the goal is to learn a domain-generalizable model while using only a small fraction of labeled data and a relatively large fraction of unlabeled data. Domain generalization (DG) methods show subpar performance under the SSDG setting whereas semi-supervised learning (SSL) methods demonstrate relatively better performance however they are considerably poor compared to the fully-supervised DG methods. Towards handling this new but challenging problem of SSDG we propose a novel method that can facilitate the generation of accurate pseudo-labels under various domain shifts. This is accomplished by retaining the domain-level specialism in the classifier during training corresponding to each source domain. Specifically we first create domain-level information vectors on the fly which are then utilized to learn a domain-aware mask for modulating the classifier’s weights. We provide a mathematical interpretation for the effect of this modulation procedure on both pseudo-labeling and model training. Our method is plug-and-play and can be readily applied to different SSL baselines for SSDG. Extensive experiments on six challenging datasets in two different SSDG settings show that our method provides visible gains over the various strong SSL-based SSDG baselines. Our code is available at github.com/DGWM.
We study the problem of subgroup discovery with Cox regression models and introduce a method for finding an interpretable subset of the data on which a Cox model is highly accurate. Our method relies on two technical innovations: the emph (Unknown sysvar: (expected prediction entropy)), a novel metric for evaluating survival models which predict a hazard function; and the emph (Unknown sysvar: (conditional rank distribution)), a statistical object which quantifies the deviation of an individual point to the distribution of survival times in an existing subgroup. Because of the interpretability of the discovered subgroups, in addition to improving the predictive accuracy of the model, they can also form meaningful, data-driven patient cohorts for further study in a clinical setting.
