Posters

Probe sampling accelerates Greedy Coordinate Gradient (GCG) and other prompt optimization methods by up to 5.6x, achieving equal or better attack success rates, making LLM safety research faster and m…

Accelerating ERM for data-driven algorithm design using output-sensitive techniques achieves computationally efficient learning by scaling with the actual number of pieces in the dual loss function, n…

Boost LLM inference speed by 3x! This paper refines blockwise parallel decoding (BPD) by cleverly refining draft predictions, resulting in faster text generation for large language models.

Boost Vision Transformer pretraining speed by 4x with novel sequence compression techniques!

Accelerated learning in games achieved! FTXL algorithm exponentially speeds up convergence to Nash equilibria in finite N-person games, even under limited feedback.

VQShape: a pre-trained model uses abstracted shapes as interpretable tokens for generalizable time-series classification, achieving comparable performance to black-box models and excelling in zero-sho…

STAR framework leverages state abstraction for consistent, low-variance off-policy evaluation in reinforcement learning, outperforming existing methods.

Absorb & Escape: a novel post-training sampling method that overcomes single model limitations by combining Autoregressive (AR) and Diffusion Models (DMs), generating high-quality heterogeneous genomi…

Transformers exhibit abrupt learning: training loss plateaus, then suddenly drops. This study uses matrix completion to demonstrate this phenomenon, providing insights into the model’s algorithmic sh…

This paper presents efficient algorithms for abductive reasoning in Logical Credal Networks (LCNs), addressing the MAP and Marginal MAP inference tasks to enable scalable solutions for complex real-wo…

A2PO: A novel offline RL method tackles constraint conflicts in mixed-quality datasets by disentangling behavior policies with a conditional VAE and optimizing advantage-aware constraints, achieving s…

Hadamard-derived Linear Binding (HLB): A novel, efficient vector symbolic architecture surpassing existing methods in classical AI tasks and deep learning applications.

ATAC-Diff: A versatile diffusion model for high-quality single-cell ATAC-seq data generation and analysis, surpassing state-of-the-art.

A single model tackles diverse audiovisual generation tasks using a novel Mixture of Noise Levels approach, resulting in temporally consistent and high-quality outputs.

This paper reveals a universal square-root growth rate for H-consistency bounds of smooth surrogate losses in classification, significantly advancing our understanding of loss function selection.

A novel post-processing framework, based on a d-dimensional generalization of the Neyman-Pearson lemma, optimally solves multi-objective learn-to-defer problems under various constraints, improving co…

New normative framework for decision confidence models diverse tasks by incorporating rewards, priors, and uncertainty, outperforming existing methods.

Unified pessimism principle in offline RL conquers data sparsity & model mismatch, achieving near-optimal performance across various divergence models.

UniSeg3D: One model to rule them all! This unified framework masters six 3D segmentation tasks (panoptic, semantic, instance, interactive, referring, and open-vocabulary) simultaneously, outperforming…

A unified confidence sequence (CS) construction for generalized linear models (GLMs) achieves state-of-the-art regret bounds for contextual bandits, notably a poly(S)-free regret for logistic bandits.