VIALS
Multimodal · 2026-08-21
Handshake AI's VIALS (Visual Interpretation of Artifacts in the Life Sciences) is a visual question-answering benchmark of 161 expert-written tasks on artifacts from biotech and pharmaceutical R&D: gels and blots, phylogenetic trees, flow-cytometry plots, plasmid maps, cell-counting images, protein-structure and binding views, and small-molecule structures. Models see only the image(s) and question at their provider's default inference settings, and a GPT-5-mini judge decides whether the final answer matches the held-out ground truth (or falls in an accepted numeric range). Accuracy is averaged over three rollouts per task; failed or refused requests count as incorrect.
Top models (higher is better)
| Model | Score |
|---|---|
| GPT-6 Astra | 49.9 |
| GPT-6.1 Sol | 49.1 |
| Gemini 4 Argon | 47.2 |
| Claude Sonnet 5.5 | 41.0 |
| Claude Opus 5.5 | 40.2 |
| Muse Wonder | 31.5 |
| Gemini 3.8 Flash | 29.4 |
| Gemini 3.7 Flash | 26.5 |
| GPT-5.6 Sol | 26.5 |
| Claude Opus 5 | 24.4 |
| Muse Spark 1.2 | 22.6 |
| Gemini 3.1 Pro Preview | 21.5 |
| Kimi K3 | 20.9 |
| Grok 4.6 | 20.5 |
| Grok 4.7 | 18.2 |
| DeepSeek V4.1 Flash | 12.6 |
| GLM-4.6V | 10.8 |
| MiniMax M3 | 10.8 |
| Mistral Medium 3.5 | 7.9 |