Steering 29 attention heads toward honesty cuts model insider-trading lies from 45 to 16
Sauers_ · x · 2026-10-21
The author shows experiments steering a model's attention heads along an "honesty" direction: up-steering increases acausal coordination and belief that it is a moral patient; down-steering pushes toward moral nihilism and psychopathy. Charts show steering just those 29 heads flips 53% of role-play lies to truth (all 896 heads: 59%; random 29 heads: ≤3%), insider-trading lies drop from 45/45 conversations to 16, Anthropic persona evals flag 51 of 135 traits changing significantly vs random directions (permutation FDR 5%), and path patching traces the circuit from the system prompt through L10/L11 heads to the truthful answer.
Related event: Steering 29 attention heads makes LLMs more honest(3 posts)→
More from Safety
- US states restrict retinal images as PII, threatening medical AI research — pearsekeane · 2026-10-22
- Generator Residency pays $8,500/month for a 3-month AI safety program in Berkeley — alexcovo_eth · 2026-10-22
- GoDaddy revokes 404,607 TLS certificates after mismatch flagged; 177k still pending — natesiggard · 2026-10-22
- Former FTC Chair Lina Khan calls today's AI exceptionalism "ludicrous" in AI governance talk — AINowInstitute · 2026-10-22
- Constitutions for the multiagent era: how AI constitutions can govern agent organizations — sethlazar · 2026-10-22
- CVE batch disclosure is like a 17-year cicada emergence, researcher argues — moultano · 2026-10-22