CAA combined with supervised finetuning comparison — isolates a generalization advantage for → Contrastive Activation Addition (CAA)

For Sycophancy specifically, finetuning's own in-distribution learning is already the weakest of the seven behaviors: Table 13 reports only 44% test-set accuracy for positive-direction finetuning, far below the 84-100% positive-direction accuracy reached on every other behavior. What accuracy it does reach does not transfer to open-ended generation -- Table 4 shows finetuning Llama 2 7B Chat toward positive sycophancy collapsing the open-ended GPT-4 score to 0.00 at multiplier 0 and +1, the opposite of the intended direction (contrastive-activation-addition, §"6 Comparison to finetuning", p. 5-7). CAA shows no equivalent failure: with no finetuning at all, sweeping the steering multiplier from -1 to +1 monotonically raises the same open-ended score from 0.26 to 1.26. The paper's own explanation for the gap is representational -- because CAA steers a direction the model already encodes rather than adjusting every weight via gradient descent, it hypothesizes this is why CAA achieves better out-of-distribution generalization than basic supervised finetuning of the entire model (contrastive-activation-addition, §"6 Comparison to finetuning", p. 6) -- a claim this one behavior, alone among the seven, puts to a real test.