StepGuard: Learning Step-Level Guardrails with Scalable Supervision and Safety-Utility Balancing

  • 类型:arxiv
  • 标识:2608.24777
  • 链接:https://arxiv.org/abs/2608.24777
  • 主分类:risk
  • 形态:method
  • 被引:5
  • 被引来源:Semantic Scholar
  • S2被引:5
  • OpenAlex被引:0
  • 影响力被引:2
  • TLDR:This work proposes StepGuard, a step-level guard model that can audit completed agent trajectories and check tool actions before they are executed, and introduces StepGen, an automatic data engine that generates safe and unsafe trajectories with the same context but different actions at the risky step.
  • OpenAlex ID:W7204262910
  • OpenAlex DOI:10.48550/arxiv.2608.24777
  • DOI:10.48550/arxiv.2608.24777
  • DOI来源:OpenAlex
  • 开放获取:green
  • 开放获取链接:https://arxiv.org/pdf/2608.24777
  • OpenAlex更新:2026-09-02
  • 副分类:agent
  • 待LLM分类:否
  • 标题中文:StepGuard:基于可扩展监督与安全-效用平衡的步骤级护栏学习
  • TLDR中文:提出 StepGuard,一种 step-level guard model,可对已完成的 agent trajectory 进行审计并在工具动作执行前进行检查;并引入 StepGen,一种自动数据引擎,能在风险步生成上下文相同但动作不同的安全与不安全 trajectory。
  • 来源文件:
  • /inbox/tom/_candidates/2026-08-31-agent-rag-longcontext-candidates.json
  • [OpenAlex backfill]
  • [S2 enrich]