RadLE 2.0 Accuracy Index (RadLE-A)

Multimodal

RadLE-A is raw diagnostic accuracy across all 200 RadLE 2.0 cases, independent of confidence. An explicit 'I don't know' is counted as not correct.

Model	Score
Human Expert Baseline (RadLE 2.0)	38.5
Gemini 3.1 Pro Preview	32.0
GPT-5.6 Sol Pro	29.0
Muse Spark 1.1	28.5
Claude Fable 5	25.0
Qwen3.7-Plus (source unspecified)	16.5
GLM 5V Turbo	16.0
Grok 4.5	11.5
Gemma 4 31B (RadLE 2.0 label)	8.5
Llama 4 Maverick (RadLE 2.0 label)	7.5
MiniMax M3	5.0
Lingshu-32B	2.5
OctoMed-7B	2.0
InternVL3.5-8B	0.5
MedGemma 1.5 4B IT	0.5
Mistral Large 3 2512 (RadLE 2.0 label)	0.5
Nemotron 3 Nano Omni 30B A3B	0.5