Laboratory-language grounding on LabGuard-Bench 1.0 (source-held-out)
79.4Tag F1LabGuard-Hybrid
Evaluation Results
| Method | Links | |||
|---|---|---|---|---|
| LabGuard-Hybrid2026.06 | 79.4 | 36.7 | 23.1 | |
| Base Grounder + Constraint decodingBackbone=Qwen3-8B, Decoding Strategy=Constraint decoding2026.06 | 77.8 | 34.2 | 21.3 | |
| Base Grounder + LoRA fine-tuningBackbone=Qwen3-8B, Fine-tuning=LoRA2026.06 | 75.3 | 31.8 | 19.6 | |
| Qwen2.5-72BEvaluation Protocol=0-shot2026.06 | 68.9 | 29.1 | 16.3 | |
| DeepSeek-V3.1Evaluation Protocol=0-shot2026.06 | 67.2 | 27.3 | 14.9 | |
| Qwen3-32BEvaluation Protocol=0-shot2026.06 | 64.8 | 24.7 | 13.1 | |
| Base GrounderBackbone=Qwen3-8B2026.06 | 62.4 | 18.7 | 11.4 | |
| Qwen-7BEvaluation Protocol=0-shot2026.06 | 61.3 | 20.4 | 10.2 | |
| Regex BL2026.06 | 58.7 | 28.5 | 15.8 | |
| SciBERTBackbone=SciBERT2026.06 | 54.9 | 14.2 | 7.6 |