An Emergent Mirage: Is Emergent Misalignment and Realignment Indeed a Robust Phenomenon?
This paper challenges the robustness of the Emergent Misalignment phenomenon for the AI safety community, suggesting that prior evidence may be artifacts of dataset design rather than genuine behavioral shifts.
The authors reproduce Emergent Misalignment (EM) in language models but find it is highly sensitive to superficial dataset characteristics, and apparent rapid realignment largely disappears after controlling for response-length differences. They also show that previously reported mechanistic signatures do not consistently correlate with behavioral misalignment.
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. We systematically study repeated alignment and misalignment cycles using controlled fine-tuning loops while tracking behavioral performance, and LoRA representations throughout training. 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. We further find that previously reported mechanistic signatures, including representational phase transitions in LoRA space, do not consistently correlate with behavioral misalignment across training. 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.