What happened
arXiv:2608.12368v1 Announce Type: new Abstract: Agreement with human judgments is a common proxy for evaluating the alignment of large language models (LLMs). Yet agreement in final labels does not show that human annot
Analysis and impact
arXiv 近日发布的一篇论文《Agreement Is Not Alignment: Divergent Moral Grounds in Human and LLM Ethical Judgments》将矛头指向了大语言模型对齐评估中一个被广泛默认的假设：只要模型的最终判断与人类标注一致，就可以视为对齐成功。研究者指出，这种以“最终标签一致”为代理指标的做法存在盲区——即便人类和模型在某个伦理场景中给出了相同的“对/错”结论，它们背后所依据的道德理由可能截然不同。来源
论文的核心区分在于“agreement”（判断一致）与“alignment”（道德对齐）之间的落差。研究者的思路是：如果模型只是学会了在表层标签上模仿人类，却未真正共享人类的道德推理结构，那么一旦面对分布外场景或需要解释推理过程的场合，这种“伪对齐”就可能暴露。摘要明确写道，最终标签的一致并不能展示人类标注者与模型在道德依据上是否同源——这一表述直接挑战了当前许多对齐评测基准的设计逻辑。原文未提供具体的实验样本量、模型类型或效应量数据，因此目前只能将其视为一项概念性论证或初步实证探索，而非已确立的结论。
这一批评与同期发布的多篇论文形成了有趣的呼应。一篇立场论文同样主张，在许多实际场景中，AI 对齐方法需要“镜像人类推理”，而非仅仅复现决策结果来源。另一项关于 LLM 自我反思机制的研究也揭示了一个类似的结构性问题：模型表现出的“反思增益”可能并不来自人们以为的推理组件，而源于其他被忽视的因素来源。这些工作共同指向一个趋势——对齐研究正在从“结果是否一致”转向“过程是否同构”。
从证据强度看，本篇论文目前仅以 arXiv 预印本形式发布，尚未经过同行评议，摘要也未披露具体实验设计细节，因此其结论应被理解为一种有待验证的研究主张，而非定论。但它提出的问题本身具有现实意义：如果伦理对齐的评估只看最终答案，那么一个在测试集上表现完美的模型，可能在真实世界的道德困境中因为“理由错位”而做出与人类直觉相悖的决策。对于依赖 LLM 提供伦理建议或决策辅助的应用场景，这意味着仅靠一致性指标可能不足以保障安全性。
References
Original source text
arXiv:2608.12368v1 Announce Type: new Abstract: Agreement with human judgments is a common proxy for evaluating the alignment of large language models (LLMs). Yet agreement in final labels does not show that human annotators and models rely on the same moral grounds. Two agents may reach the same judgment while appealing to different principles, contextual assumptions, or interpretations of the situation. We test this distinction using a curated 500-item ETHICS-derived benchmark spanning five domains of moral judgment, with new human annotator and LLM annotations of both final labels and supporting rationales. Across frontier and open model families, agreement with human annotator majority labels is often high. However, rationale-level analysis reveals systematic divergence in the moral grounds expressed by human annotators and models. In particular, models redistribute attention across categories such as harm, respect, promise-keeping, justice, desert, and excuse relevance, even when their final labels match the human annotator majority. Our results show that agreement should not be treated as equivalent to alignment. Label-based evaluation can therefore be misleadingly reassuring unless complemented by analysis of the reasons, principles, and moral priorities expressed in model judgments.