🏢 Peking University

MaVEn: A novel multi-granularity hybrid visual encoding framework significantly boosts MLLM’s multi-image reasoning capabilities by combining discrete and continuous visual representations.

LSH-MoE accelerates Mixture-of-Experts training by 1.28x-2.2x via Locality-Sensitive Hashing, significantly reducing communication costs.

Long-range feedback spiking network (LoRaFB-SNet) surpasses other models in capturing dynamic and static visual cortical representations under movie stimuli, advancing our understanding of visual syst…

LM-HT SNN: A learnable multi-hierarchical threshold model dramatically improves SNN performance, achieving near-ANN accuracy through dynamic current regulation and seamless ANN-SNN conversion.

AI agents learn to balance helpfulness and self-preservation using empathy to gauge social relationships and guide reward sharing.

Self-Supervised Controllable Generation (SCG) framework achieves brain-like associative generation by using a modular autoencoder with equivariance constraints and a self-supervised pattern completion…

Key-Grid: An unsupervised 3D keypoint detector achieving state-of-the-art semantic consistency and accuracy for both rigid and deformable objects using novel grid heatmap features.

InfiNet achieves state-of-the-art results by enabling feature interaction in an infinite-dimensional space using RBF kernels, surpassing models limited to finite-dimensional interactions.

IRE framework expedites the discovery of flat minima in deep learning, enhancing generalization and convergence. By decoupling the dynamics of flat and sharp directions, IRE boosts sharpness reduction…

HonestLLM boosts LLM honesty & helpfulness by 65.3% (Llama3-8b) and 124.7% (Mistral-7b) using training-free and fine-tuning methods, establishing principles and a new dataset (HONESET) for honesty eva…

HiCoM, a novel framework, achieves high-fidelity streamable dynamic scene reconstruction by using a hierarchical coherent motion mechanism and parallel processing to significantly reduce training time…

GS-Hider: A novel framework secures 3D Gaussian Splatting by embedding messages in a coupled, secured feature attribute, enabling invisible data hiding and accurate extraction.

GraphMorph: revolutionizing tubular structure extraction by morphing predicted graphs for superior topological accuracy.

GarmentLab: A new benchmark and simulation platform tackles garment manipulation challenges by offering realistic simulations, diverse assets, and tasks bridging the sim-to-real gap for more robust AI…

Constrained sampling solved! New functional gradient flow method (CFG) efficiently samples from constrained probability distributions via a novel boundary condition for gradient flows, achieving prov…

Prompt Adversarial Tuning (PAT) defends against LLM jailbreaking by training a protective prompt prefix. PAT uses adversarial and benign prompts to optimize this prefix, significantly reducing succes…

Uni-Mol2, a groundbreaking 1.1B parameter molecular pretraining model, reveals power-law scaling in molecular representation learning, achieving significant performance improvements on downstream task…

This research introduces a novel, automated protocol translation framework for self-driving labs, tackling the challenge of converting human-readable experimental protocols into machine-interpretable …

EnOF-SNN boosts spiking neural network (SNN) accuracy by enhancing output feature representation using a novel knowledge distillation method and ReLU activation, outperforming current state-of-the-art…

EGODE, a novel framework, leverages coupled graph ODEs and an event module to accurately model continuous and instantaneous changes in rigid body dynamics, outperforming existing methods.