StepGuard: Learning Step-Level Guardrails with Scalable Supervision and Safety-Utility Balancing
StepGuard checks agent tool actions before execution, cutting attack success by 77.3% in the paper’s agent tests.
The model is trained to judge step-level actions, not just completed agent trajectories. Its StepGen engine creates paired safe and unsafe trajectories with the same context at the risky step. Balance-GRPO is used to reduce both over-blocking and missed unsafe actions. The authors report top average accuracy among open-weight guard models, comparable to GPT-5.4, with utility down 2.8 percentage points. HF Daily Papers' note
The model is trained to judge step-level actions, not just completed agent trajectories. Its StepGen engine creates paired safe and unsafe trajectories with the same context at the risky step. Balance-GRPO is used to reduce both over-blocking and missed unsafe actions. The authors report top average accuracy among open-weight guard models, comparable to GPT-5.4, with utility down 2.8 percentage points. HF Daily Papers' note
score 5