LookThere! Sparse Vision by Reinforced Selection

On September 7, 2026, the LookThere method was published on arXiv cs.LG. This study achieved a Pareto frontier breakthrough in performance and computational complexity through an end-to-end reinforcement learning framework that processes only task-related inputs without the need for auxiliary signals. The method jointly trains a shallow input selector and a deep representation extractor, enabling the model to maintain high accuracy in high-resolution, sparse recognition scenarios such as traffic signs and billiards with only 0.2% of input data. Additionally, LookThere supports cross-task and cross-model generalization applications such as ImageNet classification, ADE20K segmentation, zero-sample classification, and counting, surpassing existing state-of-the-art selection methods and providing a generalizable and scalable framework for specialized and efficient adaptive computing.