Benchmark RadarSteerBench-Work
SteerBench-Work evaluates agent steering decisions at action boundaries in workplace scenarios across seven domains, with 106 incident-anchored scenarios and evidence-reversed mirrors, scored on correct proceed/hold boundaries.
- Released
- 2026-08-12
- Readiness
- Paper only
- Primary field
- Health & Life Sciences
Why it matters
Addresses the critical pre-commit decision in long-running agents, where a single step can have significant consequences. The benchmark reveals that models tend to over-refuse authorized actions, providing valuable insight for calibrating agent behavior to avoid both unsafe actions and unnecessary delays.
Motivation
Long-running LLM agents act through tools, and a single step can send an email, merge a pull request, or wire a payment.
Primary resources
Benchmark Radar records only publicly supported details and links back to primary sources for verification.