Alignment dynamics in LLM fine-tuning

The paper proposes a computable alignment score and derives its closed-form update during fine-tuning, to uniformly explain why existing alignment is reversed by subsequent fine-tuning. The analysis decomposes the alignment update into two parts: resilience is jointly determined by the current alignment state and the narrowness of the model distribution, and the driving force depends on how well the training distribution matches the resulting conditional posterior of aligned and unaligned completions. The framework predicts a rehearsal priming effect, i.e., prior alignment leaves a latent posterior imprint that accelerates realignment upon re-exposure to related data. The authors validate alignment reversal and accelerated realignment in safety alignment, emergent misalignment, and sentiment settings, and in controlled experiments on safety alignment confirm that rebound strength varies as the distributional posterior narrows.