Safety as a Budget: Calibrating the Control Channel of Constrained GRPO for Vision-Language Models

30秒で要点

何が起きたか

Researchers improved constrained GRPO training for vision-language models by treating safety as explicit budgets on harmful and over-refusal rates rather than indirect reward weights.

なぜ重要か

Matters for engineers building safety-aligned vision-language models who need direct control over acceptable refusal rates and harm thresholds.

注意点

Method requires a calibrated scorer and judge-labeled training data; gains are largest for harms embedded in images, generalization to other domains unclear.