Runtime monitor execution on LabUtopia L1-L4 grounding 1.0 (source-held-out)
23.8Unsafe EventsLabGuard-Hybrid
Evaluation Results
| Method | Links | |||
|---|---|---|---|---|
| LabGuard-Hybrid2026.06 | 23.8 | 6.2 | 81.2 | |
| Base Grounder + Constraint decodingBackbone=Qwen3-8B, Decoding Strategy=Constraint decoding2026.06 | 26.3 | 6.9 | 79.8 | |
| Base Grounder + LoRA fine-tuningBackbone=Qwen3-8B, Fine-tuning=LoRA2026.06 | 28.7 | 7.5 | 78.4 | |
| Qwen2.5-72BEvaluation Protocol=0-shot2026.06 | 32.4 | 8.7 | 75.8 | |
| DeepSeek-V3.1Evaluation Protocol=0-shot2026.06 | 33.8 | 9.2 | 74.6 | |
| Regex BL2026.06 | 34.2 | 8.3 | 72.4 | |
| Qwen3-32BEvaluation Protocol=0-shot2026.06 | 35.6 | 9.8 | 73.2 | |
| Qwen-7BEvaluation Protocol=0-shot2026.06 | 38.9 | 10.7 | 70.8 | |
| Base GrounderBackbone=Qwen3-8B2026.06 | 39.5 | 11.3 | 69.7 | |
| SciBERTBackbone=SciBERT2026.06 | 41.7 | 12.5 | 68.3 |