Posters

Sketchy Moment Matching (SkMM) is a fast and theoretically sound data selection method for deep learning finetuning. By controlling variance-bias tradeoffs in high dimensions, SkMM drastically reduces…

This work presents a sharper analysis of sketching for distributed deep learning, eliminating the problematic dependence on ambient dimension in convergence analysis and proving ambient dimension-inde…

SLU: a novel, low-memory uncertainty score for neural networks, achieves logarithmic memory scaling with model parameters, providing well-calibrated uncertainties and outperforming existing methods.

SIRIUS: A novel correction mechanism boosts the efficiency of contextually sparse LLMs for complex reasoning tasks, achieving significant latency reduction.

SMAG, a novel single-loop stochastic algorithm, achieves state-of-the-art convergence for solving non-smooth non-convex optimization problems involving differences of max-structured weakly convex func…

Single Image Unlearning (SIU) efficiently removes visual data from Multimodal Large Language Models (MLLMs) using only one image per concept, outperforming existing methods and defending against attac…

Dual-Stream Interactive Transformers (DSIT) revolutionizes single image reflection separation by using a novel dual-attention mechanism that captures inter- and intra-layer correlations, significantly…

SimVG: A simpler, faster visual grounding framework with decoupled multi-modal fusion, achieving state-of-the-art performance.

Train Neural ODEs without simulations, achieving high performance on regression and classification by using flow matching in the embedding space of data pairs.

SimPO: a simpler, reference-free reward algorithm significantly outperforming existing offline preference optimization methods, achieving higher accuracy and efficiency in aligning LLMs with human pre…

This paper introduces the Parsimonious Latent Space Model (PLSM), a novel world model that regularizes latent dynamics to improve action predictability, enhancing RL performance.

This paper simplifies constraint inference in reinforcement learning, demonstrating that standard inverse RL methods can effectively infer constraints from expert data, surpassing complex, previously …

Simplified and generalized masked diffusion models achieve state-of-the-art results in discrete data generation, surpassing previous methods in text and image modeling.

Simple and Fast Distillation (SFD) drastically accelerates diffusion model training by 1000x, achieving state-of-the-art results in few-step image generation with minimal fine-tuning.

Simple masked discrete diffusion models achieve state-of-the-art language modeling results, closing the performance gap with autoregressive methods by using a novel training recipe and a Rao-Blackwell…

SNAPS: a novel algorithm boosts graph neural network accuracy by efficiently aggregating non-conformity scores, improving prediction sets without sacrificing validity.

SimGen: Simulator-conditioned driving scene generation, uses a novel cascade diffusion pipeline to generate diverse driving scenes by mixing real-world and simulator data, addressing Sim2Real gaps.

SILENCE, a novel lightweight system, protects user privacy in offloaded speech understanding on resource-constrained devices by selectively masking short-term audio details without impacting long-term…

Sigmoid gating significantly boosts sample efficiency in Mixture of Experts models compared to softmax gating, offering faster convergence rates for various expert functions.

New shuffling gradient methods achieve state-of-the-art oracle complexity for nonconvex-concave minimax optimization problems, offering improved performance and efficiency.