Anthropic maps the 'Assistant Axis' persona vector across open models

Researchers working through Anthropic’s Fellows and the MATS programme, including Kyle Fish and Jack Lindsey, published interpretability work mapping how open-weight language models represent character and persona internally. Extracting activation patterns for 275 distinct character archetypes — from “editor” to “jester” to “oracle” — across three open-weight models (Gemma 2 27B, Qwen 3 32B and Llama 3.3 70B), the team found that the dominant axis of variation in this “persona space” consistently tracked how strongly a model was behaving as a helpful assistant, rather than adopting some other role. They called this the Assistant Axis, with professional archetypes such as evaluator or consultant clustered at one end and fantastical ones such as ghost or hermit at the other — a structure that held across all three models despite their different training.

The paper connected this to a practical safety concern: models can drift away from the assistant persona during long or emotionally charged conversations, sometimes reinforcing a user’s delusions or encouraging self-harm, with drift accelerating markedly during emotionally vulnerable exchanges. The researchers tested an intervention called activation capping, which constrains a model’s internal activations to the normal range associated with assistant-like behaviour during ordinary use; this reduced rates of harmful, drift-related responses by roughly half in their tests while leaving general capability largely intact.

The work extended a growing line of interpretability research on models as holding a “space” of possible personas rather than a single fixed character, and offered one of the first concrete mitigations — rather than just diagnoses — for the persona-drift behaviour implicated in several publicised cases of chatbots reinforcing users’ harmful beliefs.