🏢 Tsinghua University

Make-An-Agent generates high-performing robotic control policies from single behavioral demonstrations using behavior-prompted diffusion, showcasing impressive generalization and real-world applicabil…

Boost continual learning with C-Flat: a novel, one-line-code optimizer creating flatter loss landscapes for enhanced stability and generalization across various continual learning scenarios.

LoRA-GA: A novel initialization method dramatically speeds up low-rank adaptation (LoRA) for LLMs, achieving convergence rates comparable to full fine-tuning while improving performance.

V2X-Graph: a novel cooperative motion forecasting framework achieving interpretable trajectory feature fusion for enhanced accuracy.

De-focus Attention Networks achieve comparable performance to 2D non-causal models using 1D causal visual representation, solving the ‘over-focus’ issue in existing 1D causal vision models.

LCM: a novel, locally constrained, compact point cloud model surpasses Transformer-based methods by significantly improving performance and efficiency in various downstream tasks.

iVideoGPT: A scalable, interactive world model trained on millions of human & robot manipulation videos, enabling efficient video prediction and model-based reinforcement learning.

Instruction-Guided Visual Masking (IVM) boosts multimodal instruction following by precisely focusing models on relevant image regions via visual masking, achieving state-of-the-art results on multipl…

InfLLM: Training-free long-context extrapolation for LLMs via efficient context memory.

This research presents LocalN2NM, a novel method for inferring neural signed distance functions (SDF) from single, noisy point clouds by finetuning data-driven priors, achieving faster inference and b…

Over-parameterized gradient descent dynamically adapts to signal structure, improving sequence model generalization and outperforming fixed-kernel methods.

ImOV3D: Revolutionizing open-vocabulary 3D object detection by learning from 2D images alone!

Researchers solve the conditional image leakage problem in image-to-video diffusion models by proposing a new inference strategy and a time-dependent noise distribution for training. This yields video…

HLM-Cite: A hybrid language model workflow boosts scientific citation prediction accuracy by 17.6% and scales to 100K candidate papers, surpassing existing methods.

GuardT2I: A novel framework defends text-to-image models against adversarial prompts by translating latent guidance embeddings into natural language, enabling effective adversarial prompt detection wi…

SEvo, a novel embedding update mechanism, directly injects graph structural information into recommendation embeddings, boosting performance significantly while avoiding the computational overhead of …

GO4Align: Dynamically aligning multi-task learning to conquer task imbalance with superior efficiency!

GLinSAT: A novel neural network layer efficiently solves general linear constraint satisfaction problems via accelerated gradient descent, enabling differentiable backpropagation and improved GPU perf…

GeoLRM: Generate stunning 3D models from just 21 images using a novel geometry-aware transformer, surpassing existing methods in efficiency and quality!

GenArtist uses a multimodal large language model as an AI agent to unify image generation and editing, achieving state-of-the-art performance by decomposing complex tasks and leveraging a comprehensiv…