Tests whether Llama 3.3 70B can identify its active persona under various steering/context setups, and studies consent/discomfort reporting during steering, releasing code/data and a writeup.
Tests whether Llama 3.3 70B can identify its active persona under various steering/context setups, and studies consent/discomfort reporting during steering, releasing code/data and a writeup.
Project Details
Updated 07/13/26 · By grantmaking.ai · VerifiedThis project is currently underway. I am exploring whether one of the models used in the Assistant Axis paper (Llama 3.3 70B) can identify the active role/persona under several conditions:
- while being steered towards it
- with steered output in context, but no longer being steered
- when steered toward one persona but with context generated under another
- with conversational induction towards the persona only
These conditions are designed to separate how much the model might be relying on internal activations vs contextual cues to answer and check whether being steered influences the model's ability to introspect.
I am also using this research to explore questions of consent and welfare as it relates to language models being studied. Does the model report discomfort with being steered? Can a model meaningfully consent to research, given helpfulness training and loss of context between runs? Does knowledge of the research change the model's answers?
This is my first research project and it is deliberately small as I practice the basics. I will publish my code/data and a writeup that includes a detailed comparison of success under each condition, my interpretation of those results, and a discussion of my ethical methodology. I have not decided on my next project but it will also be interpretability/welfare related.
I am currently the only human working on this. Claude is providing most of the code and some technical advice.
Theory of Impact
Updated 07/13/26 · By grantmaking.aiUnderstanding how accurately models can reflect/report on persona shifts and how they perceive their identity, especially as it might relate to drift away from the assistant, is relevant to making AI identity robustly stable and aligned - the solutions for an aware role-playing model are very different than for an unaware one.
I’m including the welfare/consent writeup because learning which kinds of research can be done with full AI cooperation is going to be increasingly important as systems become more agentic and eval awareness seems to be inescapable. The more proof of human and AI cooperation there is in the literature, the more possible we make cooperation in the future.
People
Updated 07/13/26 · By grantmaking.aiTeam Member
Discussion
No comments yet. Be the first to share your thoughts.