GPT-4
Steering Llama 2 via Contrastive Activation Addition — inherited
OpenAI's large language model (OpenAI, 2023), used by CAA both to generate the custom hallucination and refusal contrastive datasets and to rate open-ended generations for how strongly they display a targeted behavior.
Building custom datasets by hand is slow, and scoring hundreds of open-ended generations by hand is slower still, so CAA leans on an outside model to do both. That choice trades effort for a new source of uncertainty: how much do you trust one language model's judgment about another language model's behavior?
GPT-4 does two jobs in this paper: generating the Hallucination and Refusal contrastive datasets, and rating open-ended responses on a 1-10 scale for how strongly they display a targeted behavior, with the spot-check against human ratings the authors ran to make sure that trust wasn't misplaced.
A large model from OpenAI, inherited whole
GPT-4 is OpenAI's large language model, released in March 2023 as the successor to the GPT-3 family that underlies InstructGPT (OpenAI, 2023) — substantially more capable at following complex instructions, reasoning over long contexts, and producing fluent, well-structured text, though CAA treats it as a fixed external tool rather than an object of study. The paper cites it only for its usage, not its architecture or training (contrastive-activation-addition, §"1 Introduction", p. 2). Unlike every model CAA actually experiments on, GPT-4 is never steered, finetuned, or probed — it sits entirely outside the Llama 2 family CAA modifies, brought in as an already-capable instrument for two jobs the paper needs done at a scale no team of human annotators could match cheaply.
Author of two datasets, rater of every open-ended result
Where Anthropic's Advanced AI Risk evaluation dataset and Sycophancy on NLP Survey / Sycophancy on Political Typology datasets cover five of the seven target behaviors, no equivalent human-written set exists for Hallucination or Refusal, so GPT-4 generates the contrastive multiple-choice items itself — Custom hallucination contrastive dataset following Rawte et al.'s taxonomy of unprompted versus contextually-triggered hallucination, and Custom refusal contrastive dataset contrasting refusal and compliant answers to questions the model should decline (contrastive-activation-addition, §"D Generating custom refusal dataset", p. 13). GPT-4 never sees Llama 2's weights, yet its output becomes the entire contrastive basis for a vector injected directly into Llama 2 Chat's residual stream. The same paper then uses GPT-4 downstream too, rating every Open-ended generation evaluation with GPT-4 rating response 0–10 against a per-behavior rubric for how strongly it displays the targeted behavior (contrastive-activation-addition, §"4.2 Open-ended generation", p. 5). Author and judge of the same pipeline is a dual role no earlier paper in this corpus asks one model to fill: GPT-4 both writes the material used to build two of the seven steering vectors and grades how well every one of the seven, including those two, worked.