arXiv subjects

Sarah Zhao

Publications and source records attributed to Sarah Zhao.

6 recordsLinked to original sources

arXiv2026

FUSE: Ensembling Verifiers with Zero Labeled Data

Verification of model outputs is rapidly emerging as a key primitive for both training and real-world deployment of large language models (LLMs). In practice, this often involves using imperfect LLM judges and reward models since ground truth acquisition can be time-consuming and expensive. We introduce Fully Unsupervised Score Ensembling (FUSE), a method for improving verification quality by ensembling verifiers without access to ground truth correctness labels. The key idea behind FUSE is to control conditional dependencies between verifiers in a manner that improves the unsupervised performance of a class of spectral algorithms from the ensembling literature. Despite requiring zero ground truth labels, FUSE typically matches or improves upon semi-supervised alternatives in test-time scaling experiments with diverse sets of generator models, verifiers, and benchmarks. In particular, we validate our method on both conventional academic benchmarks such as GPQA Diamond and on frontier, unsaturated benchmarks such as Humanity's Last Exam and IMO Shortlist questions.

stat.ML↗

arXiv2026

@GrokSet: multi-party Human-LLM Interactions in Social Media

Large Language Models (LLMs) are increasingly deployed as active participants on public social media platforms, yet their behavior in these unconstrained social environments remains largely unstudied. Existing datasets, drawn primarily from private chat interfaces, lack the multi-party dynamics and public visibility crucial for understanding real-world performance. To address this gap, we introduce @GrokSet, a large-scale dataset of over 1 million tweets involving the @Grok LLM on X. Our analysis reveals a distinct functional shift: rather than serving as a general assistant, the LLM is frequently invoked as an authoritative arbiter in high-stakes, polarizing political debates. However, we observe a persistent engagement gap: despite this visibility, the model functions as a low-status utility, receiving significantly less social validation (likes, replies) than human peers. Finally, we find that this adversarial context exposes shallow alignment: users bypass safety filters not through complex jailbreaks, but through simple persona adoption and tone mirroring. We release @GrokSet as a critical resource for studying the intersection of AI agents and societal discourse.

cs.SI↗

arXiv2025

Imputation-Powered Inference

Modern multi-modal and multi-site data frequently suffer from blockwise missingness, where subsets of features are missing for groups of individuals, creating complex patterns that challenge standard inference methods. Existing approaches have critical limitations: complete-case analysis discards informative data and is potentially biased; doubly robust estimators for non-monotone missingness-where the missingness patterns are not nested subsets of one another-can be theoretically efficient but lack closed-form solutions and often fail to scale; and blackbox imputation can leverage partially observed data to improve efficiency but provides no inferential guarantees when misspecified. To address the limitations of these existing methods, we propose imputation-powered inference (IPI), a model-lean framework that combines the flexibility of blackbox imputation with bias correction using fully observed data, drawing on ideas from prediction-powered inference and semiparametric inference. IPI enables valid and efficient M-estimation under missing completely at random (MCAR) blockwise missingness and improves subpopulation inference under a weaker assumption we formalize as first-moment MCAR, for which we also provide practical diagnostics. Simulation studies and a clinical application demonstrate that IPI may substantially improve subpopulation efficiency relative to complete-case analysis, while maintaining statistical validity in settings where both doubly robust estimators and naive imputation fail to achieve nominal coverage.

stat.ME↗

arXiv2024

Scalable Amortized GPLVMs for Single Cell Transcriptomics Data

Dimensionality reduction is crucial for analyzing large-scale single-cell RNA-seq data. Gaussian Process Latent Variable Models (GPLVMs) offer an interpretable dimensionality reduction method, but current scalable models lack effectiveness in clustering cell types. We introduce an improved model, the amortized stochastic variational Bayesian GPLVM (BGPLVM), tailored for single-cell RNA-seq with specialized encoder, kernel, and likelihood designs. This model matches the performance of the leading single-cell variational inference (scVI) approach on synthetic and real-world COVID datasets and effectively incorporates cell-cycle and batch information to reveal more interpretable latent structures as we demonstrate on an innate immunity dataset.

stat.ML↗

arXiv2023

Geodesic Mode Connectivity

Mode connectivity is a phenomenon where trained models are connected by a path of low loss. We reframe this in the context of Information Geometry, where neural networks are studied as spaces of parameterized distributions with curved geometry. We hypothesize that shortest paths in these spaces, known as geodesics, correspond to mode-connecting paths in the loss landscape. We propose an algorithm to approximate geodesics and demonstrate that they achieve mode connectivity.

cs.LG↗

arXiv2020

Controlling the spread of COVID-19 on college campuses

This research was done during the DOMath program at Duke University from May 18 to July 10, 2020. At the time, Duke and other universities across the country were wrestling with the question of how to safely welcome students back to campus in the Fall. Because of this, our project focused on using mathematical models to evaluate strategies to suppress the spread of the virus on campus, specifically in dorms and in classrooms. For dorms, we show that giving students single rooms rather than double rooms can substantially reduce virus spread. For classrooms, we show that moving classes with size above some cutoff online can make the basic reproduction number $R_0<1$, preventing a wide spread epidemic. The cutoff will depend on the contagiousness of the disease in classrooms.

math.PR↗