🏢 Fudan University

LexVLA: A novel visual-language alignment framework learns unified lexical representations for improved interpretability and efficient cross-modal retrieval.

Partial2Global: A novel VICL framework achieving globally optimal prompt selection, significantly improving visual in-context learning across various tasks.

TeT-Splatting: a novel 3D representation enabling fast convergence, real-time rendering, and precise mesh extraction for high-fidelity 3D generation.

BIR-D tames generative diffusion models for universal blind image restoration, dynamically updating parameters to handle various complex degradations without assuming degradation model types.

LLMs struggle with downstream tasks using mismatched data. TAIA, a novel inference-time method, solves this by selectively using only attention parameters during inference after training all parameter…

SpeechAlign: Iteratively aligning speech generation models to human preferences via preference optimization, bridging distribution gaps for improved speech quality.

S2HPruner bridges the discretization gap in neural network pruning via a novel soft-to-hard distillation framework, achieving superior performance across various benchmarks without fine-tuning.

This paper proposes penalty-based methods for simple bilevel optimization, achieving (ε, εβ)-optimal solutions with improved complexity under Hölderian error bounds.

MVInpainter: Pose-free multi-view consistent inpainting bridges 2D and 3D editing by simplifying 3D editing to a multi-view 2D inpainting task.

Self-supervised generative model MSA-Generator boosts protein structure prediction accuracy by producing high-quality MSAs, especially for challenging sequences lacking homologs.

RealMotion: a novel motion forecasting framework for continuous driving that outperforms existing methods by accumulating historical scene information and sequentially refining predictions, achieving …

AVMoE: a novel parameter-efficient transfer learning approach for audio-visual learning, dynamically allocates expert models (unimodal and cross-modal adapters) based on task demands, achieving superi…

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…

Lumen: A novel LMM architecture decouples perception learning into task-agnostic and task-specific stages, enabling versatile vision-centric capabilities and surpassing existing LMM-based approaches.

Low-precision local training, surprisingly, is sufficient for accurate federated learning, significantly reducing communication and computation costs.

LiveScene: Language-embedded interactive radiance fields efficiently reconstruct and control complex scenes with multiple interactive objects, achieving state-of-the-art results.

Novel intermediate variables regularization boosts knowledge graph completion!

This paper proposes a novel framework, the locally evolving set process, to develop faster localized iterative methods for solving large-scale graph problems, achieving significant speedup over existi…

GenRec: One diffusion model to rule both video generation and recognition!

Fourier Neural Processes (FNP) revolutionizes data assimilation by enabling accurate analysis of observations with varying resolutions, improving weather forecasting and Earth system modeling.