Study Finds Forcing a Single Letter Can Suppress LLM Refusal Behavior
A new arXiv paper reports mechanistic evidence that forcing a model to begin with a specific letter can suppress refusal responses on the WMDP safety benchmark, building on prior findings that multiple-choice formats weaken safety behavior.
Read the full story at arXiv cs.CL ↗
Timeline · 1 report
- 2026-11-09 05:00 · arXiv cs.CL
Forcing a Letter Suppresses Refusal: Mechanistic Evidence from WMDP