Rewiring just 89 neurons turns Qwen's built-in owl detector into a reply trigger, matching full SFT
maxsloef · x · 2026-11-11
Researcher maxsloef shows that after finetuning Qwen, the neuron with the largest validation-loss change was the model's own "owl" detector, which training rewired to fire right before every reply — "the owl was already in Qwen, we just switched it on." Their Sparse Rewiring method freezes everything except each neuron's detector side (gate & up vectors) with an L0 sparsity penalty, so changes stay interpretable in terms of the original model. With 124 test questions, 70+ rewired neurons produce mostly good in-character replies; the sparsest run matching full SFT (89%) uses just 89 neurons out of 442,279.
More from Models
- Researcher says Anthropic dodged questions on Mythos false positive rate; 37,966 findings, 2 exploited — OwariDa · 2026-11-12
- Grok spotted building meeting "Listen to Conversation" feature with live transcript — nima_owji · 2026-11-12
- MoonMath ships uncensored 763B DeepSeek-V4.1 via closed-form weight editing, claims zero capability loss — udmrzn · 2026-11-12
- 11 LLMs battle in 1,168 Cambridge Union debates; Claude Fable 5.2 dominates at 209-14 — AymericRoucher · 2026-11-12
- Claude app widget bugs: edits invisible to Claude, mobile remote control broken — petergyang · 2026-11-12
- Allen-Zhu's new paper: Canon layers boost reasoning 2x, even random convolution weights help — teortaxesTex · 2026-11-12