CAA combined with supervised finetuning comparison — runs head to head against → Supervised fine-tuning (SFT)

The finetuning half of this comparison is a specific, small recipe: one epoch of SGD at learning rate 1e-4 on the same multiple-choice contrastive dataset used to build CAA's steering vectors (contrastive-activation-addition, §"6 Comparison to finetuning", p. 5). Run alone, it works -- the paper's Limitations section reports that this unoptimized hyperparameter choice still reaches over 90% test-set accuracy on most of the seven behavior datasets (contrastive-activation-addition, §"10 Limitations", p. 10). But composing the two levers reveals where finetuning's control runs out. For Refusal, Table 4 shows finetuning alone reaching an open-ended score of 5.34, while adding positive steering on top of that same positive-refusal finetuning lowers it to 4.42 -- the opposite of what either lever does by itself. That reversal is the paper's clearest evidence that finetuning and CAA are not just redundant paths to the same effect: stacking them can interfere as well as compound, and the interaction has to be measured per behavior rather than assumed from either lever's solo performance.