CAA combined with supervised finetuning comparison
Steering Llama 2 via Contrastive Activation Addition — introduced
An experiment comparing CAA to supervised finetuning of Llama 2 7B Chat on the same contrastive multiple-choice dataset (one epoch, SGD, learning rate 1e-4), finding CAA can steer beyond finetuning's effect for 3 of 7 behaviors and generalizes better to open-ended generation -- notably for Sycophancy, where finetuning fails to generalize out of the multiple-choice format but CAA does not.
Finetuning is the standard way to bake a behavior into a model, and it's a much heavier intervention than adding a vector at inference time, which sets a high bar for comparison. Training Llama 2 7B Chat for one epoch on the same contrastive questions used to build a CAA vector shows what that heavier approach buys, and where it falls short.
Where the two methods diverge comes into focus next: finetuning reaches high accuracy on the exact multiple-choice format it trained on, yet for three of seven behaviors CAA still pushes further, and for Sycophancy specifically, finetuning's learned behavior fails to carry over into open-ended generation while CAA's does.
A small, deliberately unoptimized finetuning recipe
The finetuning half of this comparison is a specific, minimal recipe: one epoch of SGD at learning rate 1×10⁻⁴ on Llama 2 7B Chat, trained on the exact same multiple-choice contrastive dataset used to build CAA's steering vectors, maximizing the likelihood of the positive or negative answer token (contrastive-activation-addition, §"6 Comparison to finetuning", p. 5). Left deliberately unoptimized — the paper's Limitations section notes it did not tune epochs, learning rate, or loss function — this recipe still reaches over 90% test-set accuracy on most of the seven behavior datasets, a strong enough baseline that any gap CAA opens on top of it is not explained by a weak comparison partner.
A reversal: steering on top of finetuning can subtract
Composing the two levers exposes where finetuning's control runs out, and not always in the direction either lever moves alone. For Refusal, Table 4 shows finetuning alone reaching an open-ended score of 5.34, but adding positive steering on top of that same positive-refusal finetuning lowers it to 4.42 (contrastive-activation-addition, §"6 Comparison to finetuning", p. 7) — the opposite of what either steering or finetuning does by itself. For 3 of the 7 behaviors overall, CAA can steer beyond finetuning's effect in both directions; the Refusal reversal is the paper's clearest evidence that stacking the two levers can interfere as well as compound, and has to be measured per behavior rather than assumed from either lever's solo performance.
Sycophancy: the one behavior finetuning can't get out of distribution
Sycophancy is the sharpest case. Finetuning's own in-distribution learning is already the weakest of the seven behaviors — 44% test accuracy in the positive direction versus 84–100% for every other behavior (contrastive-activation-addition, §"I Finetuning test set accuracy", p. 17) — and what accuracy it reaches does not transfer: finetuning toward positive sycophancy collapses the open-ended score to 0.00 at multiplier 0 and +1, the opposite of the intended direction. CAA shows no equivalent failure, monotonically raising the same score from 0.26 to 1.26 across the multiplier range with no finetuning at all (contrastive-activation-addition, §"6 Comparison to finetuning", p. 7). The paper's hypothesis is representational: because CAA steers a direction the model already encodes rather than adjusting every weight by gradient descent, it generalizes out of distribution better than basic Supervised fine-tuning (SFT) of the whole model — a claim this one behavior puts to a real test.