DDx-TRACE - Confidence Alignment: leaderboard

Metric: Final-turn confidence alignment (-1 to 1): probability on exact or acceptable diagnoses minus probability on unmatched diagnoses on DDx-TRACE (211 physician-adjudicated multimodal neuroradiology cases), the model starts from limited history, requests imaging in free form (at most 6 requests), updates a four-item probabilistic differential after each turn and stops with a localized diagnosis, provider default decoding; endpoints judged against case rubrics by Gemini 3 Flash at temperature 0; higher is better. Source: arxiv.org. Saturation forecast: Around January 2027. 13 models tracked.

Top models

#ModelScore
1Gemini 3 Flash0.53
2Gemini 3.1 Pro (Preview)0.5
3GPT-5.40.18
4Claude Sonnet 4.60.12
5GPT-5.4 Mini0.04
6Qwen 3.5 35B A3B0.02
7Qwen 3.5 27B-0.07
8Gemma 3 27B-0.23
9Qwen 3.5 9B-0.27
10MedGemma-27B-IT-0.27
11Qwen 3.5 4B-0.36
12Lingshu-32B-0.42
13MedGemma 1.5 4B-0.63

Interactive version: theaggregate.ai/benchmark?slug=ddx-trace-confidence-alignment · How It Works · Data refreshed daily, snapshot 2026-10-07.