Publications

Preprints

Quantifying Uncertainty in the Presence of Distribution Shifts

We propose a Bayesian framework for uncertainty estimation that explicitly accounts for covariate shifts. While conventional approaches rely on fixed priors, the key idea of our method is an adaptive prior, conditioned on both training and new covariates.This prior naturally increases uncertainty for inputs that lie far from the training distribution in regions where predictive performance is likely to degrade. To efficiently approximate the resulting posterior predictive distribution, we employ amortized variational inference. For training, we simulate a range of plausible covariate shifts using only the original dataset and construct synthetic environments by drawing small bootstrap samples from the training data.

Slavutsky, Y., & Blei, DM. (2025). "Quantifying Uncertainty in the Presence of Distribution Shifts." Arxiv Preprint.

Variational Learning of Disentangled Representations

Existing variational methods for learning disentangled representations often suffer from leakage between latent variables. In this paper, we introduce DISCoVeR, a new method based on a dual-latent architecture and dual reconstruction pathways. We propose a new max–min objective that both maximizes the data likelihood and explicitly encourages disentanglement. We prove that this objective admits a unique equilibrium and promotes clean separation of latent factors. Empirically, DISCoVeR significantly outperforms prior methods in achieving disentangled representations.

Slavutsky Y.*, Beker O.*, Blei, DM., & Dumitrascu, B. (2025). "Variational Learning of Disentangled Representations." ArXiv Preprint.
* Equally contributing authors.

Published

Class Distribution Shifts in Zero-Shot Learning: Learning Robust Representations

Class distribution shifts challenge zero-shot classifiers, which rely on representations from training classes but are tested on unseen ones. We propose a model that assumes an unknown class attribute causes the shift and present an algorithm to learn robust representations where ERM would fail.

Slavutsky, Y., & Benjamini, Y. (2024). "Class Distribution Shifts in Zero-Shot Learning: Learning Robust Representations." Conference on Neural Information Processing Systems (NeuRIPS).

CONTESTS: a Framework for Consistency Testing of Span Probabilities in Language Models

Different estimators of joint probabilities of word spans arise by marginalizing the LLM’s scores from various completion orders. If the predictions are reliable, these estimates should be consistent. In this paper, we develop a statistical framework to test this and apply it to evaluate different LLMs. We show that both Masked Language Models and autoregressive models exhibit inconsistent predictions.

Wagner, E., Slavutsky, Y., & Abend, O. (2024). "CONTESTS: a Framework for Consistency Testing of Span Probabilities in Language Models." The 2024 Conference on Empirical Methods in Natural Language Processing (EMNLP).

Yuli Slavutsky