discovered 09 Aug 2026
assistant-axis
→ View on GitHubThe Assistant Axis is a tool designed to monitor, steer, and mitigate persona drift in large language models, particularly when they deviate from a default "helpful Assistant" persona during interactions. By projecting model activations onto the Assistant Axis, it enables real-time assessment of a model's adherence to desired behaviors and offers mechanisms to cap activations, thereby preventing harmful outputs. Notable features include interactive analysis notebooks, pre-computed persona vectors for popular models, and capabilities for detailed visualization of persona trajectories.