Logit-Gap Steering: A Forward-Pass Diagnostic for Alignment Robustness
The research introduces a diagnostic called the refusal–affirmation logit gap, which measures how much safety alignment helps a model refuse a prompt immediately at its first step. The authors developed Logit-Gap Steering, an efficient method to find short text suffixes that successfully close this gap. This means they found specific, in-distribution endings that bypass safety filters by making the model's initial refusal margin too small.