Posters

Lisa: a novel lazy safety alignment method safeguards LLMs against harmful fine-tuning attacks by introducing a proximal term to constrain model drift, significantly improving alignment performance.

LISA, a layerwise importance sampling method, dramatically improves memory-efficient large language model fine-tuning, outperforming existing methods while using less GPU memory.

LipFD: a novel method leverages audio-visual inconsistencies to accurately spot lip-syncing deepfakes, outperforming existing methods and introducing a high-quality dataset for future research.

LION: Linear Group RNNs conquer 3D object detection in sparse point clouds by enabling efficient long-range feature interaction, significantly outperforming transformer-based methods.

LinNet: A linear-time point cloud network achieving 10x speedup over PointNeXt, with state-of-the-art accuracy on various benchmarks.

Transformers’ in-context learning mirrors human episodic memory, with specific attention heads acting like the brain’s contextual maintenance and retrieval system.

Scaling causal language models reveals a connection between neural collapse properties, model size, and improved generalization, highlighting NC’s broader relevance to LLMs.

Linearly decompose & recompose Vision Transformers to create diverse-scale models efficiently, reducing computational costs & improving flexibility for various applications.

LinUProp: Linearly scalable uncertainty quantification for graphical models, achieving higher accuracy with lower labeling budgets!

Linear transformers surprisingly learn intricate optimization algorithms, even surpassing baselines on noisy regression problems, showcasing their unexpected learning capabilities.

Unlocking Causal Structures: New algorithms identify latent causal relationships from interventions, even when multiple variables are affected simultaneously.

Causal bandits with unknown graphs and soft interventions are solved by establishing novel upper and lower regret bounds, plus a computationally efficient algorithm.

Large Language Models struggle with compositional tasks, requiring exponentially more data than expected for learning compared to learning sub-tasks individually. This paper reveals surprising sample …

Lightweight Frequency Masker significantly improves cross-domain few-shot semantic segmentation by cleverly filtering frequency components of images, thereby reducing inter-channel correlation and enh…

LE3D: Real-time HDR view synthesis from noisy RAW images is achieved using 3DGS, significantly reducing training time and improving rendering speed.

LightUnbalancedOptimalTransport: A fast, theoretically-justified solver for continuous unbalanced optimal transport problems, enabling efficient analysis of large datasets with imbalanced classes.

LG-VQ: A novel language-guided codebook learning framework boosts multi-modal performance.

LG-CAV: Train any Concept Activation Vector with Language Guidance, leverages vision-language models to train CAVs without labeled data, achieving superior accuracy and enabling state-of-the-art model…

LFME: A novel framework improves domain generalization by training multiple expert models alongside a target model, using logit regularization for enhanced performance.

Lexicon3D: a first comprehensive study probing diverse visual foundation models for superior 3D scene understanding, revealing that unsupervised image models outperform others across various tasks.