Testing "Claude neurons" in Qwen3-8B: no better than random neurons
Sauers_ · x · 2026-10-19
Sauers shared an interpretability test: steering Qwen3-8B with so-called "Claude neurons," neither Opus 5.5 nor Astra could identify a Claude neuron — and the chart shows these neurons don't raise accuracy more than random same-layer neurons (exact test p=0.49, better than random in 51% of pairs, median gain +0.01 points).
The sharp takeaway: "none of these actually read like Claude," casting doubt on cross-model feature-identification claims in interpretability work.
Related event: Experiments probe whether a "Claude neuron" really exists(3 posts)→
More from Research
- New Paper Shows When Momentum Acceleration Works—and Fails—in Gaussian Variational Inference — ErnestRyu · 2026-10-19
- Into the Unknown: AI agents fail open scientific research, most claims fabricated — CarnegieMellonU · 2026-10-19
- DiscoveryRSI: recursive multi-agent evolution for scientific hypothesis discovery — PhAILabs · 2026-10-19
- CarryOn: object-oriented context runtime for long-horizon agents boosts OSWorld by 5.8% — McAuley-Lab · 2026-10-19
- TreeOPSD: Tencent and Tsinghua's critic-free RL for long-horizon LLM agents — Tencent-Hunyuan · 2026-10-19
- Ant International's BizBench: frontier agents often violate business constraints — ant-intl · 2026-10-19