🏢 University of Science and Technology of China

VIDEOLLM-MOD boosts online video-language model efficiency by selectively skipping redundant vision token computations, achieving ~42% faster training and ~30% memory savings without sacrificing perfo…

TRACER, a novel robust offline RL algorithm, uses Bayesian inference to handle uncertainty from diverse data corruptions, significantly outperforming existing methods.

A novel regularized triangle-shaped neural network framework, T-Net, achieves highly accurate and scalable logic circuit generation, significantly outperforming existing methods.

NAM: a novel neuron attribution method for MLLMs, revealing modality-specific semantic knowledge and enabling multi-modal knowledge editing.

CMCD framework tackles data sparsity in cognitive diagnosis by using monotonic data augmentation to improve accuracy and fairness of diagnostic results.

Dynamic NLOS imaging gets a speed boost! New ST-Mamba method leverages temporal consistency across frames for high-resolution video reconstruction, overcoming speed limitations of traditional methods.

TabPedia: a novel large vision-language model, achieves superior visual table understanding by seamlessly integrating diverse tasks via a concept synergy mechanism and a new benchmark.

SocraticLM achieves a Socratic teaching paradigm, surpassing GPT-4 by 12%, through a novel multi-agent training pipeline and a comprehensive evaluation system.

Neural Krylov Iteration (NeurKItt) accelerates linear system solving by using a neural operator to predict invariant subspaces, drastically reducing iteration counts and computation time.

MotionGS enhances deformable 3D Gaussian splatting for dynamic scenes by using motion flow to guide deformation, significantly improving reconstruction accuracy and outperforming state-of-the-art meth…

MILP-StuDio generates high-quality mixed-integer linear programming instances by preserving crucial block structures, significantly improving learning-based solver performance.

Masked Pre-training empowers a universal, fast zero-shot image denoiser!

LoTLIP boosts language-image pre-training for superior long text understanding by cleverly integrating corner tokens and utilizing a massive dataset of 100M long-caption images.

EpoD, a novel dynamic graph learning model, significantly improves generalization via a self-prompted learning mechanism for environment inference and a structural causal model utilizing dynamic subgr…

TextGen enhances multilingual visual text generation and editing by optimizing control information using Fourier analysis and a two-stage framework, achieving state-of-the-art results.

Constraining vision-language model tuning via persistent homology ensures consistent image-text alignment, improving few-shot learning and domain generalization.

CMuST: a novel continuous multi-task spatiotemporal learning framework tackles urban data limitations by enabling cross-interactions and task-level cooperation for enhanced generalization and adaptabi…

Two-stage recommender systems using tree structures achieve better generalization with more branches and harmonized training data distributions across stages.

FreqMark: Robust invisible image watermarking via latent frequency space optimization, resisting regeneration attacks and achieving >90% bit accuracy with high image quality.

Frolic: A label-free framework boosts zero-shot vision model accuracy by learning prompt distributions and correcting label bias, achieving state-of-the-art performance across multiple datasets.