🏢 University of Washington

Visual SKETCHPAD empowers multimodal language models (LLMs) with visual reasoning abilities by allowing them to generate intermediate sketches. This innovative framework substantially enhances LLM per…

This study disentangles best practices for learning from preference feedback in LLMs, revealing that data quality, algorithm choice, and reward model significantly impact performance.

This paper introduces the Uniform Last-Iterate (ULI) guarantee, a novel metric for evaluating reinforcement learning algorithms that considers both cumulative and instantaneous performance. Unlike ex…

Repeated self-distillation significantly reduces excess risk in linear regression, achieving up to a ’d’ factor improvement over single-step methods.

Gradient EM for over-parameterized Gaussian Mixture Models globally converges with a sublinear rate, solving a longstanding open problem in machine learning.

Using real images retrieved from a generator’s training data outperforms using synthetic images generated by that same model for image classification.

Data balancing in foundation models surprisingly reduces variance, improving model training and performance.

VATT: Text-guided video-to-audio generation, enabling refined audio control via text prompts and improved compatibility.

Swift Sampler (SS) automates the learning of efficient data samplers for deep learning, achieving significant performance gains (e.g., 1.5% on ImageNet) with minimal computational cost using only 10 p…

Generate multiple text drafts from a single language model pass with Superposed Decoding, significantly boosting efficiency!

Smoothed Energy Guidance (SEG) improves unconditional image generation by reducing self-attention’s energy curvature, leading to higher-quality outputs with fewer artifacts.

Self-Calibrating Conformal Prediction (SC-CP) marries model calibration and conformal prediction for more efficient and interpretable prediction intervals with prediction-conditional validity.

Massive language models improve with bigger datastores at inference time. A 1.4 trillion-token datastore, MASSIVEDS, shows that retrieval-based LMs outperform larger, solely-trained models on knowled…

This paper reveals that estimating only policy differences, while effective in bandits, is insufficient for tabular reinforcement learning. However, it introduces a novel algorithm achieving near-opti…

QUEST, a novel Metropolis-Hastings sampling method, generates high-quality & diverse machine translations by using quality metrics as energy functions, overcoming limitations of likelihood-based and r…

Researchers developed a query-based attack that generates adversarial prompts, fooling language models into producing harmful outputs with significantly higher success rates than previous methods, eff…

VPL: a novel multimodal RLHF personalizes AI by inferring user-specific latent preferences, enabling accurate reward modeling and improved policy alignment for diverse populations.

A novel algorithm, TIE, optimally teaches a family of linear behavior cloning learners, achieving instance-optimal teaching dimension while providing efficient approximation for larger action spaces.

Lanczos-FA, a simple algorithm for approximating matrix functions, surprisingly outperforms newer methods; this paper proves its near-optimality for rational functions, explaining its practical succes…

This research establishes the first minimax optimal algorithm for submodular maximization with bandit feedback, achieving a regret bound matching the lower bound.