Tag
该论文提出 Belief-Trajectory Energy (BTE)，通过测量输入在 Transformer 各层引起的预测修正来刻画输入特征，可用于任务难度估计、人机评论检测与生成模型归类，最高达 0.998 macro-AUROC。
This paper investigates how LLM-generated difficulty ratings for math items align with actual student performance, finding that LLMs systematically underestimate difficulty for items driven by learner misconceptions, a phenomenon termed the 'Easy Trap'.