新闻
Safety as a Budget: Calibrating the Control Channel of Constrained GRPO for Vision-Language Models
arXiv cs.AI · 发布于 · 阅读约3分钟
30秒读懂
- 发生了什么
- 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.
- 为何重要
- Matters for engineers building safety-aligned vision-language models who need direct control over acceptable refusal rates and harm thresholds.
- 注意
- Method requires a calibrated scorer and judge-labeled training data; gains are largest for harms embedded in images, generalization to other domains unclear.
- language model
- reinforcement learning
- grpo
这条新闻背后的模式
每个模式都讲清楚技术如何运作、何时值得投入，以及在哪里会失效。
The Agent Architect
每周一个模式、一个权衡、一个生产事故案例。为构建智能体系统的人准备的每周简报。
每周一封邮件,一键退订。您的地址仅用于发送简报。