Open models as subjects, frontier models as instruments
Every experiment in this theme quietly assumes a division of labor: some models are the ones being manipulated and measured, and others exist only to generate material or grade it, never touched themselves. Naming that split matters because conflating the two roles would make an experiment's results hard to trust.
The page lays out which models sit on which side, two mid-sized open models whose weights and activations are the actual subject of every steering and finetuning experiment, and a pair of frontier models that write training data, elicitation questions, and scoring rubrics without ever being probed themselves, plus a large real-conversation dataset that supplies the raw material both roles work from. It reads this split as the same subject-versus-instrument logic the corpus has used since its earliest testbeds, now built directly into the pipeline.
The paper's experimental stack divides its models into two roles that never overlap. Qwen2.5-7B-Instruct and Llama-3.1-8B-Instruct, two open-source instruction-tuned chat models around 7-8B parameters, are the experimental subjects throughout: every persona-vector extraction, steering, monitoring, and finetuning experiment reads their activations and writes to their weights. Claude Sonnet, in its role as pipeline and data-generation tool, and GPT-4.1-mini never have their own activations touched; instead Claude 3.7 Sonnet writes the trait artifacts and the trait-eliciting and EM-like dataset responses, Claude 4.0 Sonnet writes the interpolated system-prompt sequences used in the monitoring experiments, and GPT-4.1-mini serves as the trait expression score judge and, in the data-screening appendices, as an LLM-based data filter. LMSYS-Chat-1M sits on the subject side of the divide despite being a dataset rather than a model: sourced from 25 different LLMs' real conversations, it supplies the real-world validation data that gets projected onto persona vectors and used to finetune the open subject models, and separately trains the sparse autoencoders used to decompose those vectors. This division of labor is the corpus's testbed lens arriving at industrial scale: where 2017's experiments ran on MuJoCo and Atari and 2023's steering was judged by GPT-4 on Llama 2, 2025 formalizes the same split -- small open models as the manipulable subject, frontier closed models as the generation and judging instrument -- as the default architecture of the experiment itself.