🏢 Zhejiang University

ZipCache: Efficient KV cache quantization for LLMs using salient token identification, achieving 4.98x compression with minimal accuracy loss!

WISE, a novel dual-memory architecture, solves the impossible triangle of reliability, generalization, and locality in lifelong LLM editing by employing a side memory for knowledge updates and a route…

Energy-based Navigation Policy (ENP) revolutionizes Vision-Language Navigation by modeling joint state-action distributions, achieving superior performance across diverse benchmarks.

DiffewS: a novel framework leverages diffusion models for few-shot semantic segmentation, significantly outperforming existing methods in multiple settings.

UniIF: A unified model revolutionizes molecule inverse folding, achieving state-of-the-art results across protein, RNA, and material design by employing a novel geometric block attention network.

Unified generative-discriminative training boosts multimodal large language models (MLLMs)! Sugar, a novel approach, leverages dynamic sequence alignment and a triple kernel to enhance global and fin…

UniKE: A unified multimodal editing method achieves superior reliability, generality, and locality by disentangling knowledge into semantic and truthfulness spaces, enabling enhanced collaboration bet…

TopoFR enhances face recognition by aligning topological structures between input and latent spaces. Using persistent homology, it preserves crucial data structure info, overcoming overfitting. A har…

TOPA: Extending LLMs for video understanding using only text data.

TFGDA: Leveraging graph topology and feature alignment for superior semi-supervised domain adaptation.

Spec-Gaussian enhances 3D Gaussian splatting by using anisotropic spherical Gaussians for view-dependent appearance modeling, achieving superior real-time rendering of scenes with specular and anisotr…

SDEPO, a new natural policy gradient algorithm, efficiently solves zero-sum Markov games with continuous state spaces, achieving near-optimal convergence independent of state space cardinality.

Simple and Fast Distillation (SFD) drastically accelerates diffusion model training by 1000x, achieving state-of-the-art results in few-step image generation with minimal fine-tuning.

SDSGG outperforms leading scene graph generation methods by using LLMs to create scene-specific descriptions, adapting to diverse visual relations.

NewImp boosts diffusion models’ missing data imputation by curbing sample diversity and eliminating data masking, achieving superior accuracy.

Pairwise Softmax Loss (PSL) improves recommendation accuracy by enhancing Softmax Loss (SL) with alternative activation functions, resulting in tighter ranking metric surrogates and better noise resis…

PowerPM: A foundation model revolutionizing power system analysis by mastering complex ETS data through a novel self-supervised pre-training approach, achieving state-of-the-art performance.

PointAD: a novel zero-shot 3D anomaly detection method using CLIP’s strong generalization abilities to identify anomalies in unseen objects by transferring knowledge from both points and pixels.

PhyloGen uses a genomic language model to generate and optimize phylogenetic trees, offering faster and more accurate evolutionary analysis than traditional methods.

PaMoRL framework boosts model-based reinforcement learning speed by parallelizing model and policy learning stages over sequence length, maintaining high sample efficiency.