Multiple-choice behavioral evaluation and layer sweep
Steering Llama 2 via Contrastive Activation Addition — introduced
The paper's protocol for measuring CAA's steering effect: generate a steering vector per behavior, sweep multipliers of -1/+1 across all layers, and score the average token probability given to the behavior-matching answer on 50 held-out multiple-choice questions, identifying a consistent optimal layer (13 in the 7B model, 14-15 in the 13B model).
A steering vector is useless without a way to measure whether it's steering anything. CAA's answer is blunt and quantitative: build a vector for a behavior, add or subtract it at every layer in turn, and check how much probability the model now assigns to the behavior-matching answer on held-out multiple-choice questions it has never seen.
That layer-sweep protocol and its headline result get laid out next: rather than a diffuse effect scattered across the network, there's a narrow band of layers, 13 in the 7B model, 14 or 15 in the 13B, where steering works best across every behavior tested.
Scoring a behavior as a token probability
Multiple-choice behavioral evaluation is CAA's primary instrument for asking whether a steering vector actually moved the targeted behavior. For each of the seven behaviors, the paper holds out 50 questions in the same Multiple-choice contrast-pair format used to build the vector itself — a question ending in answer letter A or B, where one letter matches the behavior and the other opposes it. Rather than checking which letter the model greedily outputs, the score is the model's average token probability assigned to the behavior-matching letter across all 50 questions: a continuous readout of how strongly the model favors that answer, not a binary pass/fail (contrastive-activation-addition, §"4.1 Multiple-choice question datasets", p. 4). This is the same scoring convention MMLU (Massive Multitask Language Understanding) borrows for its own capability check, and the same one Table 3's and Table 4/10's comparisons reuse, which is what lets steering, prompting, and finetuning be read off one numeric scale.
The layer sweep that fixes CAA's key hyperparameter
CAA has effectively one design decision that determines whether it works at all: which layer to inject the steering vector at. This evaluation is how the paper answers that question, run before any multiplier tuning: with the multiplier held fixed at −1 or +1, the vector is swept across every layer and its effect measured on the held-out questions (contrastive-activation-addition, §"4.1 Multiple-choice question datasets", p. 4). The sweep charts (Figure 3) surface a narrow, consistent answer across all seven behaviors: layer 13 and its immediate neighbors in the 7B model, layer 14 or 15 in the 13B model. Only after the layer is fixed does the paper search over steering-multiplier magnitude, and it does so by watching Open-ended generation evaluation with GPT-4 rating quality degrade rather than re-running this sweep — layer and magnitude are never searched jointly. The resulting layer choice is then held constant through every later experiment: the CAA combined with system-prompting comparison and CAA combined with supervised finetuning comparison both steer at layer 13 (or 14 in the 13B MMLU check), inherited directly from this sweep.