🏢 National University of Singapore

Monomial-NFNs boost neural network efficiency by leveraging scaling/sign-flipping symmetries, resulting in fewer trainable parameters and competitive performance.

MomentumSMoE boosts Sparse Mixture of Experts’ (SMoE) performance by integrating momentum, resulting in more stable training and robust models.

MixEval revolutionizes LLM benchmarking by blending real-world user queries with existing datasets, creating a cost-effective, unbiased, and dynamic evaluation method.

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.

Localized Zeroth-Order Prompt Optimization (ZOPO) efficiently finds high-performing local optima for prompt optimization in black-box LLMs, outperforming existing global optimization methods.

Learning-to-Cache (L2C) dramatically accelerates diffusion transformers by intelligently caching layer computations, achieving significant speedups with minimal performance loss.

Learn macroscopic dynamics efficiently using only partial microscopic force computations! This novel method leverages sparsity assumptions and stochastic estimation for accurate, cost-effective modeli…

C-Procgen reveals implicit curriculum in Procgen’s multi-level training, showing learning shifts gradually from easy to hard contexts.

G-Retriever: a novel RAG approach enables conversational interaction with textual graphs, improving graph understanding and question answering efficiency while mitigating hallucination.

FreeSplat achieves state-of-the-art novel view synthesis by accurately localizing 3D Gaussians from long image sequences, overcoming limitations of prior methods confined to narrow-range interpolation…

Federated Transformer (FeT) revolutionizes multi-party fuzzy vertical federated learning by encoding fuzzy identifiers and using a transformer architecture, achieving up to 46% accuracy improvement an…

EZ-HOI: Efficient Zero-Shot HOI detection adapts Vision-Language Models (VLMs) for Human-Object Interaction (HOI) tasks using a novel prompt learning framework, achieving state-of-the-art performance …

Exo2Ego-V generates realistic egocentric videos from sparse exocentric views, significantly outperforming state-of-the-art methods on a challenging benchmark.

Optical-flow-free video semantic segmentation excels in adverse weather by merging adjacent frame information via a fusion block and a novel temporal-spatial teacher-student learning strategy.

Dynamic Tuning (DyT) significantly boosts Vision Transformer (ViT) adaptation by dynamically skipping less important tokens during inference, achieving superior performance with 71% fewer FLOPs than e…

DOGS: Distributed-Oriented Gaussian Splatting accelerates large-scale 3D reconstruction by distributing the training of 3D Gaussian Splatting models across multiple machines, achieving 6x faster train…

DETAIL: A novel attribution method reveals the impact of individual demonstrations in in-context learning, boosting interpretability and improving transformer-based model performance.

Self-supervised blind image deblurring (BID) breakthrough! A novel cross-scale consistency loss and progressive training scheme using implicit neural representations achieves superior performance wit…

Simple averaging of watermarked images reveals hidden patterns, enabling watermark removal and forgery, thus highlighting the vulnerability of content-agnostic watermarking methods.

Brain-JEPA: a novel brain dynamics foundation model leverages fMRI data via innovative gradient positioning and spatiotemporal masking to achieve state-of-the-art performance in diverse brain activity…