Optimizers Control LLM Emergent Misalignment Severity

Emergent misalignment (EM) is a critical issue in large language models where fine-tuning for a specific misaligned task can lead to broader, unintended misbehavior. Previous observations noted EM's sensitivity to training parameters, but a systematic understanding was lacking. This study conducted extensive experiments across various Qwen3 models, optimizers, datasets, and batch sizes to pinpoint the most influential factors. The findings indicate that the optimizer choice has the most substantial impact on EM severity, causing up to a seven-fold difference in misalignment rates. Surprisingly, model size within the Qwen3 family showed negligible effect, a finding corroborated across 12 models from three families using the Adam optimizer. The research also established a strong correlation between final log training loss and alignment, with optimizers explaining most residual variance. Further analysis of training dynamics showed that different optimizers traverse distinct paths in the loss-alignment space. Muon, an adaptive optimizer that best preserves alignment, was found to implicitly regularize for a more uniform distribution of singular values in the LoRA adapter. Building on this insight, the researchers successfully mitigated EM in Adam and Lion by adding a spectral regularization loss term, which encourages a flatter singular value spectrum, with minimal impact on training loss.