🏢 Tsinghua University

Zero-shot 3D scene reconstruction from single images is achieved by assembling diverse deep priors from large models, eliminating the need for 3D/2D training data and achieving superior performance.

YOCO: A decoder-decoder architecture for LLMs dramatically reduces memory usage and improves inference speed by caching key-value pairs only once.

YOLOv10: Real-time object detection achieves state-of-the-art speed and accuracy by eliminating NMS post-processing and holistically optimizing model architecture for efficiency and accuracy.

XMask3D uses cross-modal mask reasoning to achieve state-of-the-art open vocabulary 3D semantic segmentation by aligning 2D and 3D features at the mask level, resulting in precise segmentation boundar…

VisionLLM v2 unifies visual perception, understanding, and generation, excelling in various vision tasks and achieving performance comparable to task-specific models.

Latent Compression Learning (LCL) revolutionizes vision model pre-training by effectively leveraging readily available interleaved image-text data, achieving performance comparable to models trained o…

MAST: Train ultra-sparse deep MARL agents with minimal performance loss!

ProjDiff: A novel algorithm unleashes diffusion models’ denoising power for superior inverse problem solutions.

Unique3D: Single image to high-fidelity 3D mesh in 30 seconds!

UniGAD unifies multi-level graph anomaly detection, improving accuracy and zero-shot transferability by jointly modeling node, edge, and graph anomalies via a novel subgraph sampler and GraphStitch Ne…

UniAudio 1.5 uses a novel LLM-driven audio codec to enable frozen LLMs to perform various audio tasks with just a few examples, opening new avenues for efficient few-shot cross-modal learning.

Uni-Med, a novel unified medical foundation model, tackles multi-task learning challenges by using Connector-MoE to efficiently bridge modalities, achieving competitive performance across six medical …

Language model emergent abilities aren’t exclusive to large models; they emerge when pre-training loss falls below a threshold, irrespective of model or data size.

Self-Contrast Mixture-of-Experts (SCMoE) boosts MoE model reasoning by cleverly using ‘unchosen’ experts during inference. This training-free method contrasts outputs from strong and weak expert acti…

Compute-optimal protein language models are trained efficiently using scaling laws derived from a massive dataset, improving performance while optimizing compute budgets.

Train open-vocabulary 3D object detectors using only RGB images and large language models, achieving state-of-the-art performance without expensive LiDAR data.

TimeXer empowers transformers for superior time series forecasting by cleverly integrating exogenous variables, achieving state-of-the-art results on diverse benchmarks.

SuperVLAD: A new visual place recognition method boasts superior robustness and compactness, outperforming state-of-the-art techniques by significantly reducing parameters and dimensions.

ZAF: a novel replay-free continual learning method for vision-language models, significantly reduces forgetting by stabilizing zero-shot predictions.

SMART: a scalable, real-time multi-agent driving simulator using next-token prediction, achieves state-of-the-art results and zero-shot generalization.