Neuron-level GWAS + PheWAS on Qwen3-8B pinpoints safety-critical neurons
Sauers_ · x · 2026-10-16
Sauers shares a neuron-level GWAS + PheWAS analysis of Qwen3-8B, transplanting statistical genetics methods to interpretability: of 442,368 neurons scanned, 2,156 pass Bonferroni correction for association with refusing harmful requests. Neurons like L33.N3978+ are highly specific to safety behavior, safety traits show the highest inflation (λ=2.70), and neurons associated with more traits carry more similar trait sets — a systematic framework for mapping model behaviors to individual neurons.
More from Research
- Tencent squeezes 770B-parameter Hunyuan Hy4 Preview into 214 GiB at ~2.38 bits per weight — coolmum18 · 2026-10-16
- 200 chess games suggest in-context learning doesn't work yet, experimenter says — akbirthko · 2026-10-16
- Mosaic: multi-session agentic RL with hierarchical advantage estimation — LARK-Lab · 2026-10-16
- Meituan LongCat's WorldCast benchmark: GPT-Image-2.5 tops at 72.8, open-source lags — meituan-longcat · 2026-10-16
- Recursive Environment Scaling tanks six models' success rates by 57–77 points — Lingxiao Jing · 2026-10-16
- Harvard's MPT: decentralized agent swarms learn self-orchestration via bidding — Harvard · 2026-10-16