Contrastive Activation Addition (CAA) — operates on top of → Reinforcement learning from human feedback (RLHF)

CAA is never tested on a raw pretrained model. Every experiment in the paper runs on Llama 2 7B and 13B Chat, versions of Llama 2 already finetuned with RLHF for safety and instruction-following (Touvron et al., 2023) (contrastive-activation-addition, §"3 Method", p. 3). The paper treats this as a deliberate choice rather than an incidental one, framing the whole study as an investigation of how RLHF/finetuning techniques interact with activation engineering, building on top of the existing body of work on pretrained models (contrastive-activation-addition, §"1 Introduction", p. 2). That framing matters for how to read every later result: when CAA moves a behavior beyond what a system prompt or a finetuning run achieves, it is not steering a blank slate but finding a direction inside representations RLHF has already shaped. The base-versus-chat similarity analysis later in the paper is itself a way of asking what, specifically, RLHF changed. Five years after RLHF entered this corpus as a training-time alignment technique, CAA repurposes its finished output as a fixed substrate to be probed and nudged after training ends, rather than a target for retraining.