Limits of Activation-Direction Interventions for Emergent Misalignment Beyond Rank-1 LoRA

Activation-space interventions are starting to become a promising tool for controlling model behavior. This promise connects to a broader debate about actionability in interpretability: explanations are most useful when they support validated decisions or interventions, rather than merely exposing internal structure. We conduct a study on emergent misalignment (EM), a phenomenon in which fine-tuning on a narrow harmful task can lead to broadly misaligned behavior. Recent work suggests that in rank-1 model organisms, this behavior can be mediated by a small number of residual-stream directions. Adding the direction can elicit misalignment, and ablating it can suppress it. For EM specifically, this raises a hopeful possibility: if misaligned behavior is mediated by a specific direction, then interpretability may offer a direct safety intervention. We stress test this possibility under all-linear rank-32 LoRA fine-tuning, a higher-capacity PEFT setting closer to common open-model adaptation workflows than prior EM model organisms, which use rank-1 adapters restricted to a small set of layers, and ask whether this idea holds once a model has the capacity to encode behavioral change beyond a single residual-stream direction. Under rank-32 all-linear LoRA fine-tuning on a risky financial advice dataset, the largest activation and adapter-base signatures concentrate in late layers, while response-direction ablation suppresses EM most strongly in the mid-layer band. The response-derived direction remains behaviorally active, yet captures only a small fraction of the rank-32 activation shift. Finally, a matched benign-subtraction direction that appears locally effective reverses behavioral effect across nearby training configurations. Together, these results suggest that activation directions can remain useful handles on EM, but require stronger validation before being treated as reliable safety interventions in higher-capacity fine-tuning settings.