We introduce dParallel, a simple and effective method that unlocks the inherent parallelism of dLLMs for fast sampling. We identify that the key bottleneck to parallel decoding arises from the ...

In this work, we propose a new segmentation task --- reasoning segmentation. The task is designed to output a segmentation mask given a complex and implicit query text. We establish a benchmark ...

Abstract: In real classification scenarios, the number distribution of modeling samples is usually out of proportion. Most of the existing classification methods still face challenges in comprehensive ...

Abstract: Combined with the working environment, a topology optimization design plan was applied to the header frame of a combine harvester. The dynamics of the optimized header frame were evaluated ...