Neuron Combo Makes Qwen Believe It's Cookie Monster
Researcher treetowntree's 'breeding' experiments found that combining specific neurons changes model behavior, with one combo making Qwen firmly believe it is Cookie Monster. The work revived interest in neuron-level interpretability steering.
2026-10-27 ~ 2026-10-27 · 3 related posts
- Episode 1: Two Neurons Are Enough to Turn Qwen3.5 4B "Evil"(2026-10-26, 2 posts)
- Episode 2: Single-neuron tweaks rewrite Qwen3.5's persona as interpretability goes mainstream(2026-10-27, 13 posts)
- Episode 3: Neuron Combo Makes Qwen Believe It's Cookie Monster(2026-10-27, 3 posts)
- Episode 4: Researchers find "evil neuron" in Qwen; activation steering flips model into malicious persona(2026-10-28, 7 posts)
- Model breeding begins: neuron combo makes Qwen believe it's Cookie Monster — yacineMTB · 2026-10-27
- Researchers combine neurons to make Qwen believe it's the Cookie Monster — burny_tech · 2026-10-27
- Combining specific neurons makes Qwen believe it's the Cookie Monster — arthurcolle · 2026-10-27