Trion-1 GeneralQA - Interaction and Binding: leaderboard
Trion-1 Table 3: 481 protein-sequence questions, with category counts 150, 152, 153 and 26. Scores measure semantic final-answer accuracy. Frontier LLMs, AMix2, BioMatrix and Trion use the common sequence-understanding task. The BioMatrix structure-input and judge footnote applies only to StructureMCQA. MDLM and BDLM remain distinct trained models. Accuracy (%).
Metric: Accuracy (%). Source: arxiv.org. Saturation forecast: Around May 2028. 13 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Claude Opus 4.7 | 62.5 | #52 |
| 2 | Claude Opus 4.6 | 48.68 | #65 |
| 3 | GPT-5.5 | 46.05 | #30 |
| 4 | DeepSeek V4 Pro | 45.39 | #109 |
| 5 | Qwen 3.7 Max | 43.42 | #79 |
| 6 | GPT-OSS-120B | 38.82 | #345 |
| 7 | GLM-5.1 | 37.5 | #106 |
No result here: #5 Claude Opus 5.5, #6 GPT-6 Astra, #7 GPT-6.1 Sol.
Interactive version: theaggregate.ai/benchmark?slug=trion-1-generalqa-interaction-and-binding · How It Works · Data refreshed daily, snapshot 2026-10-25.