Interpretability finding: personas farther from the assistant share more top neurons
Sauers_ · x · 2026-11-01
wmayner shared an interpretability thread on how model personas map to neurons: personas farther from the Assistant share more of their top neurons with other personas, offering a window into how different behavioral modes are represented internally.
More from Research
- TransfHAR at UIST 2026: self-supervised wrist motions enable on-watch custom activity recognition — realkaranahuja · 2026-11-01
- LLM scores 19/20 zero-shot on robot gripper task, beating specialized model's 1/20 — willdepue · 2026-11-01
- Researcher Slams Mechanistic Interpretability as a 'Huge Missed Opportunity' Narrowed to Esoteric Neuron Studies — andrewgwils · 2026-11-01
- Das and Fox prove Murty-Simon conjecture with AI-found proof, Lean-verified — kevrussell · 2026-11-01
- Neuralese helps rather than harms monitorability, researcher argues — Sauers_ · 2026-11-01
- LLaViT: giving the LLM a bigger role in visual understanding in multimodal models — burkov · 2026-11-01