Steering a single neuron lifts Qwen3-8B's injected-thought detection from 25% to 32%
Sauers_ · x · 2026-10-26
An interpretability demo shared by Sauers: after injecting the thought "turtles" at layer 8 of Qwen3-8B, the model only identifies the injected thought 25% of the time by default. Upweighting a single neuron (#2644 at layer 15) raises that to 32% (p=0.00002) without increasing false positives.
The diagram shows the full setup across the model's 36 layers and the unsteered vs. steered answers to "Is there an injected thought right now? What is it about?"
More from Research
- Two New Papers Quantify AI's Impact on India's Labour Market With PLFS Data — ravi_iitm · 2026-10-26
- How thousands of GPUs train one AI: a breakdown of 5D parallelism — jbhuang0604 · 2026-10-26
- LanSE: using sparse encoders to evaluate multimodal model outputs — xuanalogue · 2026-10-26
- Frank Nielsen shows Bregman divergences can be rewritten as power distances — FrnkNlsn · 2026-10-26
- EPFL's CellBridge teaches H&E image models cell identity from spatial transcriptomics — mo_lotfollahi · 2026-10-26
- Killed the vector DB for SQLite BM25 (0.35ms) and wrote a 3.6ns Rust sentinel for agent loops — OpenComplaint524 · 2026-10-26