Three levers on one model: prompting, finetuning, steering
A chat model's behavior can already be pushed around before generation even starts, by writing a better instruction into the system prompt, or well before that, by finetuning the weights themselves. So a fair question about any new steering method is not just whether it works, but whether it earns its place next to tools people already reach for, and, later in this corpus, whether it can be folded into those tools rather than only compared against them.
The page walks through both comparisons in the corpus's earlier paper, steering layered on top of system prompts and steering set against finetuning, where it matches or beats finetuning on several behaviors and generalizes better to open-ended writing, then follows a later paper that dissolves the distinction: steering the same vector into a model while it finetunes, or folding the same instruction into every training example, so the lever once compared against training becomes part of training itself.
Contrastive Activation Addition is not offered as a replacement for the two standard levers on a chat model's behavior; the paper tests it against both and finds it composes with each. System-prompting, a custom instruction prepended to the context that Llama 2 Chat is trained to adapt to, is the first lever, chosen over few-shot prompting, which supplies prior examples of the desired behavior instead of an instruction and which the authors' preliminary experiments found less effective at steering the tested behaviors. The system-prompt comparison layers CAA on top of positive and negative system prompts and finds it can shift behavior beyond what either prompt alone achieves, and the reverse holds too: system-prompting adds control on top of a fixed steering multiplier. Supervised finetuning, training the model's weights directly on the same contrastive multiple-choice dataset used to build steering vectors, is the second lever. The finetuning comparison finds CAA can steer beyond finetuning's effect for three of seven behaviors and, more strikingly, generalizes to open-ended generation where finetuning sometimes cannot: for Sycophancy, finetuning on multiple-choice data fails to transfer to the free-text setting, while CAA transfers in every case. The computational efficiency comparison explains part of why steering is attractive as a lever of its own: generating a CAA vector needs only forward passes and takes under five minutes on one GPU, against roughly ten minutes on two GPUs to finetune the same model on a thousand examples. The 2025 persona-vectors paper collapses the distinction between these three levers rather than only comparing them side by side. Preventative steering applies CAA-style intervention not after training but during it: the model is proactively steered toward an undesired persona direction while finetuning proceeds, relieving it of the pressure to encode that trait in its own weights and so canceling out the training objective's pull toward the trait before it takes hold, a technique that preserves general capability on MMLU better than steering applied post-hoc and, applied across multiple layers, can suppress trait acquisition to near-baseline even on data designed to elicit it, without blocking the finetuning's intended domain-specific learning. Preventative prompting runs the same idea through the first lever instead of the third: a system prompt eliciting the undesired trait is prepended to every training example, reducing the model's own need to update its weights toward that trait, and performs comparably to preventative steering at a moderate coefficient, though steering keeps a wider, more flexible control range. Where the 2023 paper's comparisons asked whether steering could match or beat prompting and finetuning as separate alternatives, the 2025 paper's preventative pair asks what steering and prompting can do once fused into finetuning itself.