🏢 Tsinghua University

AdaptIR: A novel parameter-efficient method for generalized image restoration using a heterogeneous Mixture-of-Experts (MoE) architecture that achieves superior performance and generalization.

This paper presents a novel two-level online control algorithm that learns to control continuous-time linear systems under adversarial disturbances, achieving sublinear regret.

OneDet3D: A universal 3D object detector trained jointly on diverse indoor/outdoor datasets, achieving one-for-all performance across domains and categories.

High-dimensional spectral algorithms show saturation effects: Kernel Ridge Regression underperforms optimal algorithms like gradient flow when regression functions are very smooth.

Standard initialization in neural networks negatively impacts generalization ability under Neural Tangent Kernel theory, contradicting real-world performance, urging the development of improved theore…

Offline RL agents often fail in real-world scenarios due to unseen test states. SCAS, a novel method, simultaneously corrects OOD states to high-value, in-distribution states and suppresses risky OOD …

ODGEN: Boosting object detection accuracy by generating high-quality synthetic images using diffusion models conditioned on bounding boxes and text descriptions.

RHO-1, a novel language model, uses selective pretraining focusing on high-value tokens, achieving state-of-the-art results with significantly less data than existing models.

This paper introduces InfoNCA and NCA, novel frameworks for language model alignment using noise contrastive estimation, enabling direct optimization from both explicit rewards and pairwise preference…

Neural SDF inference is revolutionized by dynamically aligning 3D Gaussians to a neural SDF’s zero-level set, enabling accurate, smooth 3D surface reconstruction.

Neural-RDM: A novel framework for deep, scalable vision generation using residual diffusion models, achieving state-of-the-art results on image and video benchmarks.

Neural Collapse-inspired Feature Alignment (NCFAL) significantly boosts out-of-distribution generalization by aligning semantic features to a simplex ETF, even without environment labels.

MutaPLM: a novel protein language model, provides human-understandable mutation explanations and designs novel mutations with desirable properties using a unique protein delta network and chain-of-tho…

MultiPull: a novel method reconstructing detailed 3D surfaces from raw point clouds using multi-step optimization of multi-level features, significantly improving accuracy and detail.

Hierarchical Sinkhorn Tree (HST) robustly retrieves accurate 3D point cloud correspondences using multi-scale consistency, outperforming state-of-the-art methods.

MSPE empowers Vision Transformers to handle any image resolution by cleverly optimizing patch embedding, achieving superior performance on low-resolution images and comparable results on high-resoluti…

MSAGPT: Revolutionizing protein structure prediction by generating accurate virtual MSAs from limited data, boosting prediction accuracy by up to +8.5% TM-Score!

Mesa-Extrapolation enhances LLM extrapolation using a novel weave position encoding method, boosting performance while significantly reducing memory and inference time.

SPV-MIA, a novel membership inference attack, significantly improves the accuracy of identifying training data in fine-tuned LLMs by using self-prompt calibration and probabilistic variation.

MambaTalk: Efficient holistic gesture synthesis using selective state space models to overcome computational complexity and improve gesture quality.