Posters

MemoryFormer drastically cuts large language model computation by replacing fully-connected layers with memory-efficient hashing, enabling faster and more scalable AI.

Online Subspace Descent: a novel memory-efficient LLM training algorithm guaranteed to converge, closing the performance gap with full-rank methods.

FG²U: a novel memory-efficient algorithm for unbiased stochastic approximation of meta-gradients in large-scale bi-level optimization, showing superior performance across diverse tasks.

MeMo: a novel framework for pretraining meaningful, modular robot controllers via noise injection, enabling efficient transfer learning across different robot morphologies and tasks.

First benchmark for detecting training data in large vision-language models (VLLMs) improves data security.

SPV-MIA, a novel membership inference attack, significantly improves the accuracy of identifying training data in fine-tuned LLMs by using self-prompt calibration and probabilistic variation.

MeLLoC: Mechanism Learning for Lossless Compression, a novel approach that combines high-order mechanism learning with classical encoding, significantly improves lossless compression for scientific da…

MEGALODON: A new neural architecture for LLMs, enabling unlimited context length with improved efficiency and accuracy.

MEDIQ benchmark revolutionizes LLM evaluation by shifting from static to interactive clinical reasoning, revealing LLMs’ struggles with proactive information-seeking and highlighting the importance of…

Medformer: A novel multi-granularity patching transformer achieves state-of-the-art performance in medical time series classification, excelling in challenging subject-independent settings.

Med-Real2Sim uses physics-informed self-supervised learning to build non-invasive medical digital twins, enabling in-silico clinical trials and unsupervised disease detection.

New metrics and p-annealing improve sparse autoencoder training for better language model interpretability.

New scalable method measures per-unit interpretability in vision DNNs without human evaluation, revealing anti-correlation between model performance and interpretability.

MADPO, a novel MARL framework, uses mutual policy divergence maximization with conditional Cauchy-Schwarz divergence to enhance exploration and agent heterogeneity in sequential updating, outperformin…

New method efficiently measures how well AI models memorize training data, revealing that open-source models memorize less than expected.

Boosting LLMs’ abstract reasoning via ‘Meaningful Learning’: A new dataset and learning paradigm significantly enhance LLMs’ capacity for abstract reasoning, moving beyond simple memorization.

Researchers establish basis-free conditions for SGD learnability in two-layer neural networks learning subspace-sparse polynomials with Gaussian input, offering insights into training dynamics.

MC-DiT: A novel training paradigm for masked diffusion models achieving state-of-the-art image generation by leveraging clean-to-clean reconstruction.

MEow, a novel MaxEnt RL framework, achieves superior performance by unifying policy evaluation and improvement steps, enabling exact soft value function calculation without Monte Carlo approximation.

Optimizing against learning agents: New algorithms and computational limits revealed!