Tag
该论文提出了一个以用户为中心的审计框架，用于区分基准内容的贡献者（源用户）与评估结果所代表的群体（目标用户），并在16个评估资源上发现美国中心化惯例占据主导、文档中普遍存在源-目标用户语境差距，且评分指标会显著影响性能衡量。
This paper tests whether same-named emotion labels (e.g., fear, surprise) mean the same thing across GoEmotions, XED, and BRIGHTER datasets, finding large scale mismatches, reduced cross-corpus transfer performance, and cases where classifiers give opposite conclusions depending on training resource.
This paper analyzes hallucination in large language models as a structural consequence of three architectural decisions: self-attention's co-occurrence learning, maximum likelihood estimation training objective, and autoregressive decoding's left-to-right commitment. It maps each mechanism to specific hallucination types and argues that dataset pathologies amplify but do not cause these vulnerabilities.