Posters

GNAN: a novel interpretable graph neural network achieving accuracy comparable to black-box models.

Hybrid Preference Optimization (HyPO) outperforms existing offline methods for fine-tuning LLMs by leveraging both offline and online data, achieving better performance and efficiency.

ESCAIP, a novel neural network architecture, dramatically boosts the speed and accuracy of atomic simulations by leveraging attention mechanisms, enabling efficient large-scale modeling across diverse…

Leveraging data heterogeneity, this study reveals that standard SGD implicitly learns invariant features across multiple environments, achieving robust generalization without explicit regularization.

Deep learning models’ success hinges on understanding gradient descent’s implicit bias. This study reveals how this bias influences layer collaboration, revealing a decreasing trend in adversarial rob…

Researchers extended implicit bias theory to multiclass classification using a novel framework, proving that gradient descent prefers simple solutions even with complex alternatives.

Adam’s implicit bias revealed: On separable data, Adam converges towards the maximum l∞-margin solution, a finding contrasting with gradient descent’s l2-margin preference. This polynomial-time conver…

LoRA’s initialization significantly impacts finetuning; initializing matrix A randomly and B to zero yields better performance than vice-versa due to enabling larger learning rates.

Lowering a neural network’s geometric complexity during pre-training enhances neural collapse and improves transfer learning, especially in few-shot scenarios.

Researchers developed a framework for analyzing stochastic adaptive learning rate algorithms, providing exact risk and learning rate curves, revealing the importance of data covariance and uncovering …

Finetuning’s impact on worst-group accuracy is surprisingly nuanced, with common class-balancing methods sometimes hurting performance; a novel mixture method consistently outperforms others.

R3GAN, a minimalist GAN baseline, surpasses state-of-the-art models by using a novel regularized relativistic GAN loss and modern architectures, proving GANs can be trained efficiently without relying…

New research precisely defines the computational limits of training large language models, revealing a sharp threshold based on parameter matrix entries, paving the way for faster algorithms.

New ‘Feature Speed Formula’ predicts & controls deep learning’s hierarchical feature learning by linking hyperparameter tuning to the angle between feature updates and backward pass.

Novel algorithms trace the optimal balance between clustering quality and fairness, revealing all non-dominated solutions for various objectives.

Large language models (LLMs) struggle with factual inconsistencies (‘hallucinations’) and the ‘reversal curse,’ where information recall depends heavily on the input order. This work reframes the cur…

State-space models (SSMs) rival transformers in language modeling, but their capabilities remain unclear; this paper rigorously analyzes SSM expressivity, revealing unique strengths and limitations, i…

Transformers learn to perform in-context learning of Markov chains hierarchically, progressing from simpler unigram strategies to more complex bigram solutions, with the presence of simpler solutions …

Breaking neural network parameter symmetries leads to faster training, better generalization, and improved loss landscape behavior, as demonstrated by novel asymmetric network architectures.

Offline model-based RL methods fail as dynamics models improve; this paper reveals the ’edge-of-reach’ problem causing this and introduces RAVL, a simple solution ensuring robust performance.