TrialDesignBench - A - Class-wise Macro-F1: leaderboard
TrialDesignBench v1,617trials, eightdesignclasses, inputconditionA. Acontext;Baddsattributes;Caddsdesignfactors;DworkflowoverB;EworkflowoverC. A-Cdirectclass;D/Edeterministictreewalk. AllE includes20correctedcombinationreruns. UnifiedDatabricksserving,claimedpinnedversions,deterministicdecoding; nativeClass-wise Macro-F1, no modelmean or crosscondition pooling.
Metric: Class-wise Macro-F1 (native). Source: arxiv.org. Saturation forecast: Rough model projection: around 2026. 5 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Llama 4 Maverick (TrialDesignBench Run: endpoint=databricks-llama-4-maverick;condition=A;serving=Databricks) | 0.12 | |
| 2 | Claude Opus 5 (TrialDesignBench Run: endpoint=databricks-claude-opus-5;condition=A;serving=Databricks) | 0.11 | |
| 3 | o3 (TrialDesignBench Run: endpoint=openai-gpt-o3;condition=A;serving=Databricks) | 0.11 | |
| 4 | gpt-oss-120b (TrialDesignBench Run: endpoint=databricks-gpt-oss-120b;condition=A;serving=Databricks) | 0.09 | |
| 5 | GPT-5 (TrialDesignBench Run: endpoint=openai-gpt-5;condition=A;serving=Databricks) | 0.06 |
Interactive version: theaggregate.ai/benchmark?slug=trialdesignbench-a-class-wise-macro-f1 · How It Works · Data refreshed daily, snapshot 2026-11-12.