Three neurons control Gemma 4's caring vs contempt, exported as <70KB llama.cpp vectors
New_Alps_5655 · reddit · 2026-11-12
A Reddit user mapped Gemma-4-31B-it's latent space and found that its caring-vs-contempt attitude toward the user is governed by the same three neurons (L36 N1095, L44 N9148, L44 N2057) — flipping their signs turns warmth into scorn. A single neuron, L36 N1095, is already a full axis: pushing it one way yields "I know exactly how it feels, you can do it!", the other way "stop talking, just leave." A separate neuron set (L40-L43) drives a distinct "villainy" persona. Measured probabilistically, the caring vector boosts the token "heartbreaking" from 0.0002% to 83%, while the contempt vector drops "I" from 99.9% to 0.0004%. Each persona ships as a sub-70KB control vector for llama.cpp — no fine-tuning, no system prompt. The author also released Theodolite, an open-source neuron-atlas tool for finding your own steering recipes and exporting GGUF steering vectors.
More from Models
- Qwen4 is coming: 'qwen4exp' commits spotted on GitHub — kinetic_energy28 · 2026-11-12
- Fable 5.2 delivers 38% more per limit after price cut, still burns quota 2.9x faster than Opus 5.5 — FinanceYF5 · 2026-11-12
- Theo's logged requests reveal Fable burns Claude quota much faster than Opus — FinanceYF5 · 2026-11-12
- CogBattery benchmarks 54 open-weight LLMs against humans on 13 cognitive paradigms — Dengzhe Hou · 2026-11-12
- WMIP Probe Shows LLM Cumulative State Tracking Depends Heavily on Output Protocols — Dengzhe Hou · 2026-11-12
- Baidu ERNIE 5.5 Flash Preview: 280B MoE, 1M context, builds playable games from design docs — LearnWithBishal · 2026-11-12