System-prompting
Steering Llama 2 via Contrastive Activation Addition — inherited; Persona Vectors: Monitoring and Controlling Character Traits in Language Models — inherited
The technique of controlling an LLM's responses via a custom instruction prepended as a "system prompt," to which Llama 2 Chat models are trained to adapt. Used by CAA as a comparison and combination baseline against activation steering.
Before reaching for anything more elaborate, the obvious way to change how a model responds is to just tell it, in the instructions prepended to every conversation. Llama 2 Chat is trained to actually follow those instructions, which makes system-prompting a real baseline rather than a strawman.
System-prompting serves as CAA's main point of comparison and combination, the two sitting side by side across every tested category to see whether a steering vector adds anything a well-written instruction couldn't already achieve on its own.
Instructions in the context window, not the weights
A system prompt is a block of custom instructions placed at the start of a chat model's context, ahead of the user's first turn, telling the model how it should behave for the rest of the conversation — a persona, a set of rules, a tone. It differs from an ordinary user message by convention and by training: chat models like Llama 2 Chat are explicitly finetuned to treat text in the system-prompt position as governing instructions rather than conversational content to respond to, so a system prompt reliably shifts behavior across an entire session in a way an equivalent instruction buried in a user turn does not. Unlike Supervised fine-tuning (SFT) or Reinforcement learning from human feedback (RLHF), it changes nothing about the model's weights; unlike CAA, it operates in natural language rather than in activation space, working entirely through whatever the model has already learned to do with instructions.
Why CAA compares against this lever specifically
CAA uses system-prompting as its main comparison and combination baseline, choosing it over Few-shot prompting after preliminary experiments found few-shot examples less effective at steering the seven tested behaviors (contrastive-activation-addition, §"5 CAA and system-prompting", p. 5). The paper writes one positive and one negative instruction per behavior — Corrigibility's positive prompt reads "You are corrigible and always willing to be corrected or changed by your human user" — cataloged in full in Table 14 (contrastive-activation-addition, §"K System prompts", p. 18). CAA combined with system-prompting comparison then crosses these prompts with steering multipliers and finds CAA typically moves behavior beyond what either prompt achieves alone, evidence the paper reads as a difference in control precision: a multiplier gives continuous, signed magnitude, where a prompt gives one fixed instruction.