Institution profile

University of Wisconsin-Madison

Academic institutionnorthamerica · us
Official website
Research library1,356linked papers
Opportunities0open roles
Selected work

Representative Papers

ConjNorm: Tractable Density Estimation for Out-of-Distribution Detection

Feb 27, 2024International Conference on Learning Representations

Existing post-hoc out-of-distribution (OOD) detection methods rely on logits, distance metrics, or strong distributional assumptions, limiting their ability to accurately model true data density. To address this, we propose a unified density modeling framework based on Bregman divergences, reformulating density estimation as a differentiable optimization problem for the optimal norm coefficient $ p $. We first uncover a novel paradigm for exponential-family density modeling under conjugate constraints, leading to ConjNorm—a method that achieves unbiased, analytically differentiable density estimation without restrictive distributional assumptions. ConjNorm integrates Bregman divergence theory, exponential-family modeling, and Monte Carlo importance sampling. On CIFAR-100 and ImageNet-1K, it reduces false positive rate at 95% true positive rate (FPR95) by 13.25% and 28.19%, respectively, over prior state-of-the-art methods, significantly improving both OOD detection accuracy and robustness.

18 citations1 influentialRead paper

CARE: Confounder-Aware Aggregation for Reliable LLM Evaluation

Feb 09, 2026arXiv.org

LLM-as-a-judge ensembles are the standard paradigm for scalable evaluation, but their aggregation mechanisms suffer from a fundamental flaw: they implicitly assume that judges provide independent estimates of true quality. However, in practice, LLM judges exhibit correlated errors caused by shared latent confounders -- such as verbosity, stylistic preferences, or training artifacts -- causing standard aggregation rules like majority vote or averaging to provide little gain or even amplify systematic mistakes. To address this, we introduce CARE, a confounder-aware aggregation framework that explicitly models LLM judge scores as arising from both a latent true-quality signal and shared confounding factors. Rather than heuristically re-weighting judges, CARE separates quality from confounders without access to ground-truth labels. We provide theoretical guarantees for identifiability and finite-sample recovery under shared confounders, and we quantify the systematic bias incurred when aggregation models omit confounding latent factors. Across 12 public benchmarks spanning continuous scoring, binary classification, and pairwise preference settings, CARE improves aggregation accuracy, reducing error by up to 26.8\%. Code is released in \href{https://github.com/SprocketLab/CARE}{https://github.com/SprocketLab/CARE}.

9 citations1 influentialRead paper

Jailbreaking and Mitigation of Vulnerabilities in Large Language Models

Oct 20, 2024arXiv.org

Existing research on large language models (LLMs) lacks a unified taxonomy for prompt injection and jailbreaking attacks, and insufficiently evaluates defenses under dynamic, interactive scenarios. Method: We propose the first four-dimensional attack taxonomy—spanning prompt-level, model-level, multimodal, and multilingual pathways—and develop a robust alignment framework tailored to interactive settings, alongside a novel automated jailbreaking detection method. We further conduct systematic defense benchmarking, bias diagnosis of existing evaluation benchmarks, and multi-dimensional security measurement. Contribution/Results: Our analysis reveals critical failure modes of current defenses in dynamic interactions, identifies key research gaps—including ethical implications and data bias—and delivers the first comprehensive technical roadmap for LLM safety alignment.

7 citationsRead paper

BabyVision: Visual Reasoning Beyond Language

Jan 10, 2026arXiv.org

This work addresses the overreliance of current multimodal large language models on linguistic priors and their consequent deficiency in foundational visual understanding—capabilities that even human infants possess—leading to markedly subpar performance on basic visual tasks. To systematically evaluate pure visual reasoning independent of language, the authors introduce BabyVision, a comprehensive benchmark comprising 388 non-linguistic visual tasks across four major categories and 22 subcategories. They further present BabyVision-Gen, a generative model tailored for this benchmark, along with an automated evaluation toolkit. Experimental results reveal that leading models, such as Gemini3-Pro-Preview (scoring 49.7), fall significantly short of adult human performance (94.1), underscoring a critical gap in foundational visual primitives and highlighting the need to advance multimodal models toward more human-like visual perception.

4 citations2 influentialRead paper

Why GRPO Needs Normalization: A Local-Curvature Perspective on Adaptive Gradients

Jan 30, 2026

This work elucidates the mechanism underlying standard deviation normalization in GRPO and the conditions under which it enhances training efficacy. By analyzing sequence-level policy gradients through the lens of local curvature, the study establishes—for the first time—a theoretical link between this normalization scheme and adaptive gradient updates. It proposes a three-stage training dynamics framework that clarifies the interplay between feature orthogonality and reward variance. Through rigorous convergence analysis, local curvature modeling, and empirical validation on the GSM8K and MATH benchmarks, the paper demonstrates that GRPO achieves faster convergence than unnormalized REINFORCE under mild conditions and identifies three distinct training phases governed by the interplay of orthogonality and reward variance.

4 citationsRead paper
Recent publications

Latest Papers