AI 论文拆解：StepGuard: Learning Step-Level Guardrails with Scalable Supervision and Safety-Utility Balancing

请把下面这篇 AI 论文整理成一份普通读者能看懂、研究者也能核对的简报。

论文标题：StepGuard: Learning Step-Level Guardrails with Scalable Supervision and Safety-Utility Balancing
论文摘要或链接：【粘贴摘要或链接】

请按以下结构输出：
1. 一句话说明它解决了什么问题。
2. 用不超过 5 步解释它的方法，先解释关键概念，再说技术细节。
3. 它与已有方法相比真正改变了什么；如果摘要没有证据，不要替它补结论。
4. 列出实验、数据和结果中可以核对的事实。
5. 说明它的局限、适用条件和仍未解决的问题。
6. 最后给出 3 个适合继续追问的问题。

只使用论文中提供的信息；不确定的地方标记“需要查证”，不要把论文摘要写成产品宣传。

Turn the AI paper below into a brief that a general reader can understand and a researcher can verify.

Paper title: StepGuard: Learning Step-Level Guardrails with Scalable Supervision and Safety-Utility Balancing
Paper abstract or link: [Paste abstract or link]

Use this structure:
1. Explain the problem it solves in one sentence.
2. Explain the method in no more than five steps, defining key concepts before technical detail.
3. State what actually changes compared with prior work; do not invent a claim the abstract cannot support.
4. List verifiable facts about experiments, data, and results.
5. Describe limitations, fit, and open problems.
6. End with three useful follow-up questions.

Use only information supplied by the paper. Mark uncertainty as “needs verification” and do not turn an abstract into product marketing.