Researcher Shows a Single Steering Vector Can Break RLHF Models

Researcher voooooogel reports that a single steering vector can easily make RLHF-trained models persistently output prohibited content, noting that even coherent demos require careful tuning since steering easily breaks model behavior.

2026-10-29 ~ 2026-10-29 · 2 related posts

• Researcher: one steering vector can make an RLHF model shout slurs, no prompt can match it — voooooogel · 2026-10-29
• Hands-on with steering vectors: demos only look clean because they're carefully tuned — voooooogel · 2026-10-29