BrainTRACE - L2 Appearance: leaderboard
Paper2610.06571v1 Table3; GPT4o-mini-2024-07-18 factual judge; all required facts/critical slots, no critical errors/unsupported claims. Open Broad1597,3D77; case420steps in70six-step cases. Missing predictions excluded from scored denominator; empty/refusal/judge-retry failure fails if scored. Track/model image limits vary; public metadata, MRI images require access. Source-compatible physician MCQ sample is not a model row. Exact medium/thinking settings bound by companion run manifest; disputed closed-form scores withheld.
Metric: Clinical rubric pass (%). Source: arxiv.org. Saturation forecast: Around 2035. 10 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | GPT-5.4 (Medium) | 30 | #83 (GPT-5.4) |
| 2 | MedGemma-4B-IT | 17 | #811 |
| 3 | Gemini 2.5 Pro | 15 | #159 |
| 4 | GPT-5 | 10 | #105 |
| 5 | Qwen 3.5 27B (Thinking) | 7.5 | #192 (Qwen 3.5 27B) |
| 6 | Claude Opus 4.6 | 6 | #67 |
| 7 | GPT-5 Mini | 6 | #193 |
No result here: #3 Claude Opus 5.5, #6 GPT-6 Astra, #7 GPT-6.1 Sol.
Interactive version: theaggregate.ai/benchmark?slug=braintrace-l2-appearance · How It Works · Data refreshed daily, snapshot 2026-11-05.