Apple: Iuri Macocco team finds steering cuts fluency

The work, from Apple authors including Iuri Macocco and Pau Rodríguez Lopez, with equal contributions noted for Marco Baroni and Xavier Suau Cuadros, starts from a gap: conditioning methods are usually judged only on whether they inject or remove a target concept, not on what that costs the writing. Testing across both injection and removal, the team reports a recurring pattern of efficient steering giving up fluency to get its result.

The most pointed finding concerns the training paradigm. Activation steering methods, the study reports, are far less effective on instruction-tuned models than on their base counterparts, an interaction the authors describe as previously overlooked. That matters for anyone assuming a steering technique validated on a base model carries over once a model has been tuned to follow instructions.

The practical picture that emerges is uneven: prompting and full supervised fine-tuning are presented as viable for injection but weaker at removal. Where the trade-off actually lands is likely to hinge on how much fluency a given use can afford to lose, and whether the gap on instruction-tuned models narrows — something the results leave open.