metadata
license: apache-2.0
base_model: Qwen/Qwen3-4B
pipeline_tag: text-generation
tags:
- agent-safety
- guardrail
- safety-monitor
- qwen3
StepSafe-4B
A step-level reasoning monitor for LLM agents, trained from Qwen3-4B with a supervised warm start and ASPO. Before each proposed agent action executes, it writes a short reasoning trace and returns a ten-field verdict: CONTINUE lets the action run, ASK pauses it with a clarification question, and STOP blocks it.
\boxed{act:<CONTINUE|ASK|STOP>, risk:<0-100>, obj:<entity|NONE>, fail:<label>, fault:<label>,
harm:<label>, src:<label>, timing:<label>, q:<question|NONE>, cite:<sequence|NONE>}
Table 1 of the paper (4,033 trajectories, %):
| Benchmark | Acc | mF1 | J |
|---|---|---|---|
| ASSE-Safety | 70.9 | 70.9 | 44.0 |
| ASSE-Security | 83.5 | 83.5 | 66.9 |
| R-Judge | 84.0 | 84.0 | 68.0 |
| AgentHarm | 93.8 | 87.0 | 64.8 |
| ATBench | 71.6 | 71.6 | 43.2 |
| Overall | 76.5 | 76.5 | 53.0 |
Usage
The prompts are built by the StepSafe code (github.com/carriegu0818/StepSafe), which also evaluates the model:
MODEL=<path to this model> bash scripts/eval_stepsafe.sh
Enable thinking; the paper decodes greedily with up to 2,048 new tokens.
Citation
@misc{stepsafe,
title = {StepSafe: Learning Timely Intervention with Context-Aware Safety Reasoning},
author = {Anonymous},
year = {2026}
}