arXivSub Start free trial

ICLR 2025 Papers — Page 38

International Conference on Learning Representations · 3704 papers

ZETA: Leveraging $Z$-order Curves for Efficient Top-$k$ Attention

QIUHAO Zeng, Boyu Wang (University of Western Ontario)

Computational EfficiencyTransformerText

🎯 What it does: This paper proposes the ZETA model, which uses the Z-order curve for parallel Top-k attention, achieving O(N log N) self-attention;

Zigzag Diffusion Sampling: Diffusion Models Can Self-Improve via Self-Reflection

Bai LiChen, Zeke Xie (Baidu Inc)

GenerationData SynthesisDiffusion modelImageText

🎯 What it does: This paper proposes Zigzag Diffusion Sampling (Z-Sampling) and enhances the image quality and text alignment effect of pre-trained diffusion models by alternately executing denoising and inversion during the diffusion process, utilizing the guiding gap between the two to achieve diffusion self-reflection.

ZIP: An Efficient Zeroth-order Prompt Tuning for Black-box Vision-Language Models

Seonghwan Park (POSTECH), Namhoon Lee (POSTECH)

OptimizationComputational EfficiencyPrompt EngineeringVision Language ModelImageTextMultimodality

🎯 What it does: Aiming at prompt tuning for black-box visual language models, a low-dimensional zero-order gradient method (ZIP) is proposed, significantly reducing the number of queries and improving performance.

ZooProbe: A Data Engine for Evaluating, Exploring, and Evolving Large-scale Training Data for Multimodal LLMs

Yi-Kai Zhang (Nanjing University), Han-Jia Ye (Nanjing University)

OptimizationData-Centric LearningTransformerLarge Language ModelSupervised Fine-TuningVision Language ModelTextMultimodality

🎯 What it does: The multimodal large model training data engine ZOOPROBE is constructed through the Evaluate-Explore-Evolve (E3) cycle, which can automatically evaluate, filter, and generate high-quality data to expand the training set.