Exploration-Lab/dim-discovery-archive

该数据集是NeurIPS 2025论文《Geometry of Decision Making in Language Models》的官方实现/发布，研究了大型语言模型（LLMs）在多选问答（MCQA）设置中通过内在维度（ID）分析隐藏表示的几何特性。研究包括28个开放权重的Transformer模型，估计了各层的ID，并量化了MCQA任务的每层性能。数据集提供了额外的实验结果、内在维度估计和逐层性能分析，为研究人员在该领域的进一步探索提供了有价值的资源。This dataset is the official implementation/release for the NeurIPS 2025 paper Geometry of Decision Making in Language Models, which studies the internal decision-making processes of large language models (LLMs) through the lens of intrinsic dimension (ID) in a multiple-choice question answering (MCQA) setting. The research includes 28 open-weight transformer models, estimates ID across layers, and quantifies per-layer performance on MCQA tasks. The dataset provides additional experimental results, intrinsic-dimension estimations, and layer-wise performance analysis, serving as a valuable resource for researchers for further exploration in this area.