🏢 University of Wisconsin-Madison

Yo’LLaVA personalizes Large Multimodal Models (LMMs) to converse about specific subjects using just a few images, embedding concepts into latent tokens for efficient and effective personalized convers…

TSDS: A novel framework selects optimal training data for efficient large language model finetuning using only a few examples, boosting performance.

Calibrated robust fine-tuning boosts vision-language model accuracy and confidence in out-of-distribution scenarios by using a constrained multimodal contrastive loss and self-distillation.

Shuffled SGD’s convergence is now better understood through a primal-dual analysis, yielding tighter bounds that align with its superior empirical performance.

Alchemist, a novel automated labeling system, reduces data annotation costs by 500x compared to LLMs while improving accuracy by an average of 12.9%.

PSPS: a novel task-agnostic framework enables valid and efficient ML-assisted statistical inference for virtually any task, simply using summary statistics from existing analysis routines!

This paper achieves minimax-optimal bounds for learning near-optimal policies in average-reward MDPs, addressing a long-standing open problem in reinforcement learning.

First sample-efficient algorithm for LMDPs without separation assumptions, achieving near-optimal guarantees via novel off-policy evaluation.

New algorithm reliably learns Gaussian halfspaces with significantly improved sample and computational complexity compared to existing methods, offering strong computational separation from standard a…

Colander: a novel auto-labeling technique boosts data efficiency by 60%, optimizing confidence functions for maximum coverage with minimal error.

Dope: Efficient algorithms optimize human preference elicitation for learning to rank, minimizing ranking loss and prediction error with absolute and ranking feedback models.

This work presents a computationally efficient algorithm that robustly learns a single neuron despite adversarial label noise and distributional shifts, providing provable approximation guarantees.

HaloScope leverages unlabeled LLM outputs to accurately detect AI hallucinations without human annotation, significantly outperforming existing methods.

Adaptive Sampling with Approximate Expected Futures (ASAp) ensures LLMs generate grammatically correct outputs that closely match the model’s original probability distribution.

Faster algorithms achieve optimal excess risk in user-level private stochastic convex optimization, overcoming limitations of prior methods without restrictive assumptions.

This paper presents an efficient algorithm to compute near-optimal deterministic policies for constrained reinforcement learning problems, solving a 25-year-old computational complexity challenge.

Consistency Purification boosts certified robustness by efficiently purifying noisy images using a one-step generative model, achieving state-of-the-art results while maintaining semantic alignment.

New method for eigenvector estimation achieves optimal rates without coherence dependence, improving low-rank matrix denoising and related tasks.

BackdoorAlign defends against fine-tuning-based LLM jailbreaks using a ‘backdoor trigger’ to enforce safety alignment during inference, effectively mitigating risks with minimal additional safety exam…

This paper introduces a unified ITR estimation framework using covariate balancing weights, achieving significant gains in robustness and effectiveness compared to existing methods.