🏢 Peking University

RectifID personalizes image generation by cleverly guiding a diffusion model using off-the-shelf classifiers, achieving identity preservation without needing extra training data.

Boosting large language model performance after knowledge editing: A new method (D4S) minimizes model damage by regulating the explosive growth of parameter layers, enabling multiple effective edits.

RAGRAPH, a novel retrieval-augmented graph learning framework, boosts GNN generalization by integrating external graph data, significantly outperforming state-of-the-art methods.

This paper introduces a novel three-phase neural network framework that significantly enhances the quality of polarimetric images by complementarily fusing degraded noisy and blurry snapshots, preserv…

PrivCirNet accelerates private deep learning inference by cleverly transforming DNN weights into circulant matrices, converting matrix-vector multiplications into efficient 1D convolutions suitable fo…

MGPO: Efficient online RL adaptation via prompt optimization of pre-trained multi-goal transformers.

PiSSA, a novel parameter-efficient fine-tuning method, surpasses LoRA by initializing adapter matrices using the principal components of the original model, achieving faster convergence and enhanced p…

Researchers developed a universal adversarial patch to fool real-world large vision-language models (LVLMs) across multiple tasks, without needing access to internal model details.

Panacea: a novel LLM alignment method achieving Pareto optimality via online preference adaptation using a single model.

This paper proposes ‘generalized quasar-convexity’ to optimize problems with multiple probability distributions, offering adaptive algorithms with superior iteration complexities compared to existing …

Opponent modeling based on subgoal inference (OMG) outperforms existing methods by inferring opponent subgoals, enabling better generalization to unseen opponents in multi-agent environments.

OpenGaussian achieves 3D point-level open vocabulary understanding using 3D Gaussian Splatting by training 3D instance features with high 3D consistency, employing a two-level codebook for feature dis…

Score Implicit Matching (SIM) revolutionizes diffusion model distillation by creating high-quality, single-step generators from complex, multi-step models, achieving comparable performance and enablin…

Small-size Graph Neural Networks effectively solve Linear Programs!

OmniJARVIS: Unified vision-language-action tokenization enables open-world instruction-following agents via unified multimodal interaction data.

SeqComm, a novel multi-level communication scheme, tackles multi-agent coordination by leveraging asynchronous decision-making and a two-phase communication process for improved efficiency and theoret…

MotionBooth: A new framework enabling precise control over both object and camera movements in customized text-to-video generation, achieving high-quality video while maintaining training efficiency.

MoLE: Mixture of Low-rank Experts enhances human-centric text-to-image diffusion models by using low-rank modules trained on high-quality face and hand datasets to improve the realism of faces and han…

MO-DDN: A new benchmark and coarse-to-fine exploration agent boosts embodied AI’s ability to handle multi-object, preference-based task planning.

MemoryFormer drastically cuts large language model computation by replacing fully-connected layers with memory-efficient hashing, enabling faster and more scalable AI.