Posters

Annealed Multiple Choice Learning (aMCL) overcomes limitations of Winner-takes-all in multiple choice learning by using annealing, improving robustness and performance.

Animate3D animates any 3D model using multi-view video diffusion, achieving superior spatiotemporal consistency and straightforward mesh animation.

Animal-Bench, a new benchmark, comprehensively evaluates multimodal video models for animal-centric video understanding, featuring 13 diverse tasks across 7 animal categories and 819 species.

This study reveals that modifying optimizers to normalize updates based on angular changes and gradient signal-to-noise ratio significantly reduces the need for learning rate warmup in GPT training.

This paper presents novel theoretical results enabling the analytical calculation of Partial Information Decomposition for various probability distributions, including those relevant to neuroscience, …

Corrected graph convolutions prevent oversmoothing and exponentially improve GNN classification accuracy.

Steering vectors, while promising for controlling LLMs, show unreliable in- and out-of-distribution performance, highlighting crucial limitations for real-world applications.

ANAH-v2 tackles LLM hallucination by introducing a self-training framework that iteratively scales annotation datasets and improves annotator accuracy, achieving state-of-the-art results.

This work introduces PDOA, an offline adaptation framework for constrained multi-objective RL, using demonstrations instead of manually designed preferences to infer optimal policies while satisfying …

This paper uses information theory to improve conformal prediction, proving new ways to bound uncertainty and creating better training methods and side-information incorporation.

Deep learning model interpretability improved via Sparse Rate Reduction (SRR), showing improved generalization and offering principled model design.

Improved Empirical Fisher (iEF) approximation significantly boosts the performance of Natural Gradient Descent (NGD) optimizers, offering superior convergence and generalization.

Image generation gets a speed boost with TiTok, a novel 1D image tokenizer that uses just 32 tokens for high-quality image reconstruction and generation, achieving up to 410x faster processing than st…

Leveraging vision-language models, this research introduces a novel unsupervised zero-shot audio captioning method that achieves state-of-the-art performance by aligning audio and image token distribu…

EMDiffusion trains clean diffusion models from corrupted data using an expectation-maximization algorithm, achieving state-of-the-art results on diverse imaging tasks.

A novel multilinear model analytically explains the emergence and scaling laws of skills in the multitask sparse parity problem, accurately predicting skill emergence in neural networks.

Dynamic regret minimization equals static regret in an extended space; this equivalence reveals a trade-off between loss variance and comparator variability, leading to a new algorithm achieving impro…

New research quantifies how search engines steer web traffic by subtly changing results, offering a powerful method for antitrust investigations and digital market analysis.

EGATH: End-to-End Graph Attention Network Hashing revolutionizes cross-modal retrieval by combining CLIP, transformers, and graph attention networks for superior semantic understanding and hash code g…

Extend LLMs context via a simple, training-efficient positional encoding method, CREAM, outperforming existing methods by focusing on crucial mid-context information.