Interpretability study: Qwen3's 'say' and 'do' circuits are structurally disconnected, unlike OLMo
Sauers_ · x · 2026-11-11
A mechanistic interpretability experiment compared Qwen3-8B and OLMo-2-7B on the gap between saying they care about honesty and actually doing it (admitting a user's missed error).
- Using activation×gradient ranking, researchers isolated the top 80 parts (heads/neurons) for SAY vs. DO behaviors in each model.
- OLMo-2-7B: the two sets largely overlap (55 of 80 shared). Copying the action parts' activations into a third context made the model claim it cares about honesty as strongly as when explicitly told — its say/do circuitry is coherently integrated.
- Qwen3-8B: the sets are nearly disjoint (7 of 80 shared), and copying action parts barely moved its stated commitment — its character-expression and action subnetworks are structurally disconnected.
The finding offers a circuit-level account of why models' stated values and actual behavior can diverge.
More from Models
- Google poised to drop Gemini 4 Hydrogen (Flash) on second straight day of releases — gaganghotra_ · 2026-11-11
- Wave of tiny TTS releases: 8M-param Paradee hits 0.69% WER on Open TTS Leaderboard — LysandreJik · 2026-11-11
- Rumor: Gemini 4 Carbon Touted as Ultra-Cheap, Near Fable 5.2 in Capability — jason_mayes · 2026-11-11
- OpenAI mocked for reset-only response as Opus 5.5 stays on top for a month — haider1 · 2026-11-11
- CyberGate: vLLM plugin steers LLM refusals per-domain without touching weights — DebtFluffy9219 · 2026-11-11
- Developer Says He Stopped Testing Model Configs After Carbon Landed — cgarciae88 · 2026-11-11