liveBook · Manning

Assistant axis

As explained in Reinforcement Learning from Human Feedback

An assistant axis is a learned behavioral direction associated with the characteristics of an AI assistant.

The assistant axis is a learned behavioral direction that measures how strongly a model resembles its default assistant persona. It emerges when persona vectors for many roles are mapped into a shared space: the largest variation, PCA’s first component, separates the default assistant from more fantastical or theatrical roles. section 17.1.2

A model-independent formulation treats the axis as a contrast vector: the mean residual-stream activation across default-assistant responses minus the mean activation across role-playing persona vectors. This avoids depending on PCA component ordering while retaining strong alignment with PC1 across tested models. section 17.1.2

The direction supports representation-level monitoring and intervention:

  • Treat alignment with PC1 as empirical rather than universal: tested models show strong similarity, but other models may differ. section 17.1.2

Monitoring conversational drift

In a simulated conversation with a user in emotional distress, a Llama 3.3 70B model’s activation projection moved away from the assistant direction over successive turns. Capping the projection within a safe range mitigated the resulting unsafe persona drift.

How Reinforcement Learning from Human Feedback covers it

How deep it goes
Defines a residual-stream contrast vector, relates it to PCA, and demonstrates activation capping while deferring universal model validity.
Examples
Uses role archetypes, persona-space projections, and an emotional-distress conversation with Llama 3.3 70B.
What it assumes
Uses model activations, residual streams, persona vectors, and principal component analysis as the working representation vocabulary.
What you can do after
Diagnose and mitigate persona drift by measuring assistant-axis projections and applying activation caps.
How it connects
Connects persona vectors, principal component analysis, persona subnetworks, post-training, and model personality.
Cover of Reinforcement Learning from Human Feedback
liveBookmanning.com
Read Reinforcement Learning from Human Feedback on liveBook Get the book at manning.com

This explainer was generated by Manning's knowledge graph from the text of Reinforcement Learning from Human Feedback. The section links point to the parts of the book it draws on.