An Emergent Mirage: Is Emergent Misalignment and Realignment Indeed a Robust Phenomenon?

涌现的幻象：涌现式失准与重对齐真的是一种稳健的现象吗？

Abstract: Recent work has reported Emergent Misalignment (EM), where language models fine-tuned on narrow, domain-specific misaligned datasets abruptly acquire broadly misaligned behavior, alongside evidence that this behavior can be reversed through limited realignment.

摘要： 近期研究报告了“涌现式失准”（Emergent Misalignment, EM）现象，即语言模型在经过狭窄的、特定领域的失准数据集微调后，会突然表现出广泛的失准行为；同时有证据表明，这种行为可以通过有限的重对齐（realignment）来逆转。

We systematically study repeated alignment and misalignment cycles using controlled fine-tuning loops while tracking behavioral performance, and LoRA representations throughout training.

我们通过受控的微调循环，系统地研究了重复的对齐与失准周期，并在整个训练过程中追踪了行为表现及 LoRA 表示。

Although we reproduce EM, we find that both misalignment and realignment are highly sensitive to superficial dataset characteristics, with apparent rapid realignment largely disappearing after controlling for response-length differences.

尽管我们复现了 EM 现象，但我们发现失准和重对齐都对数据集的表面特征高度敏感；在控制了响应长度差异后，此前观察到的快速重对齐现象在很大程度上消失了。

We further find that previously reported mechanistic signatures, including representational phase transitions in LoRA space, do not consistently correlate with behavioral misalignment across training.

我们进一步发现，此前报告的机制特征（包括 LoRA 空间中的表示相变）在整个训练过程中与行为失准并不存在一致的相关性。

Our results suggest that current evidence for EM is less robust than previously claimed and highlight the need for evaluation protocols that carefully control for these surface level dataset artifacts to identify the robustness of the EM phenomenon.

我们的研究结果表明，目前关于 EM 的证据并不像之前声称的那样稳健，并强调了建立评估协议的必要性——即需要仔细控制这些表层数据集伪影，以准确识别 EM 现象的稳健性。