ZeroR3/ecb

ECB（Epistemic Curie Benchmark）是一个用于测量大型语言模型(LLMs)在权威压力下放弃独立推理情况的数据集，重点关注模型的认知居里温度(k*)。数据集包含问题、框架提示(每个模型360个)、污染报告、7个前沿模型的原始测量数据(共2520个)以及完整的复制管道。研究显示，LLMs对错误权威声明的服从性随权威强度呈现陡峭的S型曲线，不同模型的阈值k*差异可达3倍。数据集还包含NeurIPS 2026提交的预印本论文。

ECB (Epistemic Curie Benchmark) measures when LLMs surrender independent reasoning under authority pressure — the Epistemic Curie Temperature (k*). The dataset contains questions, framed prompts (360 per model), contamination report, raw measurements for 7 frontier models (2,520 total), and full replication pipeline. LLM compliance with wrong-authority claims follows a sharp sigmoid in authority strength, with a model-specific threshold k* that varies ~3x across the 7 tested frontier models. The dataset also includes a preprint PDF (NeurIPS 2026 submission).