EVENT DOSSIER
arXiv:2609.0445v1 proposes the “fractional mechanism shift” phenomenon, indicating that when large language models (LLMs) consistently provide incorrect answers, their scoring mechanism shifts, leading to the invalidation of confidence certificates. The study found that in multi-agent systems, this shift reduces coverage from 90% after calibration to 74% (standard alpha = 0.10). In multi-choice question-answering tasks with open weights, attackers can reduce the coverage of low-confidence items from 87% to 47%, while the monitored average value remains high. Moreover, the failure of the decision layer causes systems that should have upgraded to trust the attacker’s incorrect answers enough to take action. Existing standard-compliant repair methods cannot solve this problem because the distribution of errors remains unchanged, but the scoring behavior of the models changes.
arXiv:2609.04445v1 提出“分数机制偏移”现象，即 LLM 在同伴一致给出错误答案时，其评分机制发生漂移导致置信证书失效。该研究指出，这种偏移会破坏多智能体系统中的顺从预测（Conformal Prediction），使覆盖度从校准后的 90% 降至 74%（标准 alpha = 0.10）。在开放权重的多选择问答任务中，攻击者针对低置信度项可将该子组的覆盖度从 87% 压低至 47%，而监控的平均值仍保持较高。此外，决策层失效导致本应升级的系统反而对攻击者的错误答案产生足够信心并执行行动。现有标准顺从修复方法无法解决此问题，因为问题分布未变，而是模型的评分行为发生了改变。