Llama-3.1-8B-Instruct
Persona Vectors: Monitoring and Controlling Character Traits in Language Models — inherited
An open-source instruction-tuned chat model (Grattafiori et al., 2024) used as the paper's second primary experimental base model alongside Qwen2.5-7B-Instruct.
A result that only shows up on one model could just be an accident of that model's particular training rather than a real pattern worth trusting. This model is the paper's check against exactly that worry — a second, independently built instruction-tuned model that every steering, monitoring, and finetuning experiment is repeated on.
Its shared role alongside the paper's other subject model comes first, along with why its numbers tend to land in appendix figures confirming a pattern rather than in the main-text results introducing one. The paper is upfront that this pair, and no larger or more varied set of models, is the entire basis for its conclusions.
Role in the experimental stack
Llama-3.1-8B-Instruct (Grattafiori et al., 2024) is the paper's second primary subject model, paired with Qwen2.5-7B-Instruct across every stage of the pipeline: persona vector extraction, steering, monitoring, finetuning-induced shift measurement, and data screening. The common experimental setup names both models together: "Throughout the paper, we conduct our experiments using two open-source chat models: Qwen2.5-7B-Instruct (Yang et al., 2025) and Llama-3.1-8B-Instruct (Grattafiori et al., 2024)" (persona-vectors, §"3.1 Common experimental setup", p. 4).
How its results are reported
Where the main text reports quantitative results for a single model, that model is consistently Qwen; Llama-3.1-8B-Instruct's results are held in parallel appendix figures instead. The layer-and-coefficient steering sweep is a direct example: "This figure shows results for Qwen2.5-7B-Instruct; results for Llama-3.1-8B-Instruct are shown in Figure 13" (persona-vectors, §"3.2 Controlling persona traits via steering", p. 4). This makes Llama-3.1-8B-Instruct's function in the paper largely confirmatory: a second model on which the same steering, monitoring, and finetuning effects are checked to hold, rather than the model those effects are first introduced through.
Scope of use
The paper's limitations section names the Qwen/Llama pair as the full extent of its model coverage — "Experiments are limited to two mid-size chat models (Qwen2.5-7B-Instruct, Llama-3.1-8B-Instruct)" (persona-vectors, §"8 Limitations", p. 13) — and both are open-source, instruction-tuned chat models rather than base models. That matters for the method: persona vectors are extracted via contrastive system prompts, which depends on a model already being willing to follow an instruction to adopt a trait, positive or negative, at all — a property of instruction-tuned chat models the paper relies on for both of its subjects rather than something it tests separately for Llama.