SafeRLHF Test

anchor

Findings

• IC-556All 9 LLM-based guard models exhibit significant miscalibration with average ECE exceeding 10% across 12 public benchmarks for both prompt and response classification [eval]
• IC-559Contextual calibration is most effective for prompt classification while temperature scaling is more effective for response classification, but no single post-hoc method fully resolves miscalibration [eval]