Steering 29 attention heads makes LLMs more honest
An open-source project shows that steering 29 attention heads in Qwen3.5-9B toward honesty cuts insider-trading rates from 45 to 16 and turns over half of role-play lies into truths.
2026-10-21 ~ 2026-10-22 · 3 related posts
- Steering 29 attention heads toward honesty cuts model insider-trading lies from 45 to 16 — Sauers_ · 2026-10-21
- Steering 29 attention heads turns half of Qwen3.5-9B's role-play lies into truthful answers — Sauers_ · 2026-10-21
- Researcher Steers Model Attention Heads Toward Honesty, Flips Nihilism Traits — voooooogel · 2026-10-22