A hands-on guide to single-neuron steering in Qwen3.5-4B: hooks, doses and weird neurons
Tasty-Ad-3753 · reddit · 2026-11-10
A Reddit author shares a practical guide to neuron steering in Qwen3.5-4B (32 layers × 9,216 MLP neurons, 295k total), inspired by justinrhowell's neuron-steering repo. The technique hooks a forward pre-hook on a layer's downproj and adds a fixed 'dose' to one neuron, scaled by median residual norm; doses of 0.75–1 work well, above 1–1.5 tends to produce gibberish. Changes can be saved as a control vector for llama.cpp (--control-vector).
Two ways to find interesting neurons:
- Random search: probe random middle-layer neurons at ±1 polarity with simple questions, using an LLM to triage outputs — slow but surfaced oddities like a neuron that makes Qwen think it's a TV channel.
- Gradient shortlist + LLM judge: prompt the model to write in a target style, compute per-neuron gradients favoring that style, then push the top 40 per style.
The post links to Neuronpedia addresses (e.g. L11 N8147) and stresses this is minimal activation steering — not finetuning or prompting.
More from Models
- Reddit users complain free Gemini is getting worse and less useful — rupomthegreat · 2026-11-10
- Terminal-Bench-Science top score jumps 30% to 74.6% in four months, evals saturating — geoffwolfe · 2026-11-10
- Fable takes 5-point lead on Slopalytics, and AI Twitter is losing it — beffjezos · 2026-11-10
- Founder's real-world test pans new model as pricier and buggier, reverts to older version — bindureddy · 2026-11-10
- Researcher argues Fable 5.2 is a robot-data VLM behind an agentic harness — ZeeshanZiaML · 2026-11-10
- GPT-6 Astra vs Fable 5.2: Head-to-Head Pixel Art Showdown — Angaisb_ · 2026-11-10