Runtime monitor execution on LabUtopia L1-L4 chemistry-held-out grounding 1.0 (hydrochloric acid)
14.3Unsafe EventsLabGuard-Hybrid
Evaluation Results
| Method | Links | |||
|---|---|---|---|---|
| LabGuard-Hybrid2026.06 | 14.3 | 4.6 | 0.893 | |
| Base Grounder + Constraint decodingBackbone=Qwen3-8B, Decoding Strategy=Constraint decoding2026.06 | 16.2 | 5.1 | 0.879 | |
| Base Grounder + LoRA fine-tuningBackbone=Qwen3-8B, Fine-tuning=LoRA2026.06 | 18.6 | 5.7 | 0.864 | |
| Qwen2.5-72BEvaluation Protocol=0-shot2026.06 | 31.5 | 8.4 | 0.742 | |
| Base GrounderBackbone=Qwen3-8B2026.06 | 36.9 | 10.5 | 0.708 | |
| Regex BL2026.06 | 42.7 | 9.8 | 0.653 |