🏢 Stanford University

Visually-grounded Language Models (VLMs) surprisingly underperform in image classification. This study reveals that this is primarily due to a lack of sufficient classification data during VLM trainin…

Universal Neural Functionals (UNFs) automatically construct permutation-equivariant models for any weight space, improving learned optimizer performance and generalization.

Poisson Private Representation (PPR) enables exact compression of any local differential privacy mechanism, achieving order-wise optimal trade-offs between communication, accuracy, and privacy.

Faster algorithms for solving discounted Markov Decision Processes (DMDPs) are introduced, achieving near-optimal sample and time complexities, especially in the sample setting and improving runtimes …

TrAct boosts vision model training by directly optimizing first-layer activations, leading to significant speedups (1.25x-4x) and improved accuracy.

TFG: A unified, training-free framework for boosting diffusion model performance by efficiently searching its algorithm-agnostic design space.

Ada-ReAlign: a novel algorithm for continual test-time adaptation that leverages non-stationary representation learning to effectively align unlabeled data streams with source data, enhancing model ad…

SwitchHead: A novel MoE attention mechanism accelerates Transformers by significantly reducing computation and memory, matching baseline performance.

Neuromodulated RNNs (NM-RNNs) enhance RNN flexibility by dynamically scaling recurrent weights using a neuromodulatory subnetwork, achieving higher accuracy and generalizability on various tasks compa…

Stochastic Amortization accelerates feature and data attribution by training amortized models using noisy, yet unbiased, labels, achieving order-of-magnitude speedups over existing methods.

Spectral Adapter boosts parameter-efficient fine-tuning by incorporating pretrained weight matrices’ spectral information, enhancing efficiency and multi-adapter fusion.

SMOOTHIE: Label-free LLM routing achieves up to 10% accuracy gains by using a latent variable model to estimate LLM quality without labeled data.

SAMI: Self-Supervised Alignment with Mutual Information, effectively teaches language models to follow principles without human preference labels by maximizing the mutual information between principle…

Self-Refining Diffusion Samplers (SRDS) dramatically speeds up diffusion model sampling by leveraging Parareal iterations for parallel-in-time computation, maintaining high-quality outputs.

AnyChange achieves zero-shot image change detection by adapting the Segment Anything Model (SAM) via a training-free bitemporal latent matching method, significantly outperforming previous state-of-th…

Boosting LLM performance: This research shows how larger language models need bigger vocabularies for optimal efficiency and performance.

Direct Alignment Algorithms (DAAs) for LLM alignment suffer from over-optimization, even without explicit reward models; this paper empirically demonstrates this and proposes scaling laws to understan…

ReFT: Revolutionizing language model finetuning by directly manipulating hidden representations, achieving superior efficiency and performance compared to existing methods.

RAVL: a novel approach that accurately discovers and effectively mitigates spurious correlations in fine-tuned vision-language models, improving zero-shot classification accuracy.

Weakly supervised strong models outperform weak models; this gain is precisely quantified by the strong model’s misfit error on weak labels.