Two misaligned models share 102 top neurons against 13 expected by chance
jfonsecarivera · x · 2026-11-02
For each model, the author ranked all 442,368 neurons by how differently they fire on evil vs. normal answers. The two top-419 lists share 102 neurons, versus about 13 expected by chance — strong evidence that emergent misalignment across independent runs reuses the same internal circuitry.
Related event: Independently misaligned Qwen models light up the same power-themed neurons(4 posts)→
More from Research
- Anthropic's FlashProve brings provable inference to LLMs with <100x overhead on Llama4 — AccBalanced · 2026-11-04
- Paperclip builds biomedical datasets in one shot, cutting months-long curation to about a minute — james_y_zou · 2026-11-04
- Embodied Agent Arena: Closed-Weight VLMs Lead All Five Robot Task Domains, Open-Weight Models Lag — Haojian Huang · 2026-11-04
- FPT University Paper Shows RAG Improves Stock Market Information Retrieval Accuracy and Latency — Huan Phong Thai · 2026-11-04
- Alibaba's adaptive GUI agent training cuts environment interactions by 56-61% — Yuxiang Chai · 2026-11-04
- Large-Scale Study: AI Reviewers Often Ignore How Much You Improved Your Paper — umd-zhou-lab · 2026-11-04