Safety & Grounding on M3TAVR (test)
8.1Hallucination RateR-CGA + InternVL3.5-8B
Evaluation Results
| Method | Links | ||
|---|---|---|---|
| R-CGA + InternVL3.5-8BFramework=TAVR-VLM, Backbone=InternVL3.5-8B2026.06 | 8.1 | 62.4 | |
| R-CGA + Qwen3-VL-8BFramework=TAVR-VLM, Backbone=Qwen3-VL-8B2026.06 | 8.7 | 60.3 | |
| Gemini-3 ProParadigm=Closed-Source Frontier Models2026.06 | 11.5 | — | |
| Claude-4 OpusParadigm=Closed-Source Frontier Models2026.06 | 12.9 | — | |
| GPT-4oParadigm=Closed-Source Frontier Models2026.06 | 13.8 | — | |
| InternVL3.5-8BParadigm=Open-Source Multimodal Models2026.06 | 14.7 | 51.6 | |
| Qwen3-VL-8BParadigm=Open-Source Multimodal Models2026.06 | 17.9 | 50.1 | |
| LLaVA-Med-1.5Paradigm=Open-Source Multimodal Models2026.06 | 24.2 | 45.4 | |
| METransformerParadigm=Report Generation Baselines2026.06 | 33.8 | 23.4 | |
| R2Gen-MambaParadigm=Report Generation Baselines2026.06 | 34.6 | 21 | |
| MCLN-TransformerParadigm=Report Generation Baselines2026.06 | 37.9 | 17.5 |