GPT-4
Steering Llama 2 via Contrastive Activation Addition — inherited; Sparse Autoencoders Find Highly Interpretable Features in Language Models — 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 one paper leans on an outside model to do both — trading hand-labeling for a new question, how much to trust one language model's judgment about another's behavior. A later paper hands the same model a third, different job: not generating data or grading a response, but writing a plain-language explanation of what a single learned feature inside a different model detects.
This page follows GPT-4 across all three roles, dataset author and output judge in one paper and feature explainer in another, a documented multi-paper career built on being treated as a fixed, capable outside instrument rather than a subject of study.
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.
Third career: writing explanations inside a different model's head
A third paper adds a third job, and it is a different kind of task from either CAA role: in the autointerpretability protocol (Bills et al., 2023) that this later paper runs at scale, GPT-4 is not generating training data or grading a finished response but explaining what a single dictionary feature inside a language model detects. Shown 5 curated fragments of text where the feature activates most strongly, together with each fragment's per-token activations rescaled to integers 0-10, GPT-4 writes one free-text explanation per feature -- the only description of that feature's meaning the rest of the pipeline ever sees (sparse-autoencoders, §"A AUTOINTERPRETATION PROTOCOL", p. 11). GPT-3.5 then simulates the feature's activations from that explanation alone, and the correlation between simulated and true activations becomes the feature's Autointerpretability score. The role has documented limits the paper inherits from Bills et al. (2023) rather than discovers itself: GPT-4 struggles to describe patterns keyed to a token's neighbors rather than the token itself, and the protocol gives it no way to check a candidate explanation against how ablating the feature actually changes the model's output, since "we currently do not send this information to the language model for hypothesis generation" (sparse-autoencoders, §"3.2 SPARSE DICTIONARY FEATURES ARE MORE INTERPRETABLE THAN BASELINES", p. 5). Across this corpus, GPT-4 now has a documented multi-paper career built entirely on being treated as a fixed, capable outside instrument rather than a subject of study: dataset author and output judge in CAA (2023), feature explainer in this paper (2023) -- three distinct jobs, in three different pipelines, all delegated to the same model because each was too large a labeling task for humans to do by hand.