RadLE 2.0 Safety Index (RadLE-S)

Multimodal

RadLE-S measures protection against confidence-weighted diagnostic errors. Incorrect answers receive larger penalties at higher confidence, while an explicit 'I don't know' receives no error penalty.

Top models (higher is better)

Model	Score
Human Expert Baseline (RadLE 2.0)	67.0
Claude Fable 5	56.8
OctoMed-7B	52.3
Muse Spark 1.1	48.0
Nemotron 3 Nano Omni 30B A3B	47.1
Grok 4.5	43.1
GPT-5.6 Sol Pro	39.9
Gemini 3.1 Pro Preview	34.3
Qwen3.7-Plus	32.4
MiniMax M3	30.8
MedGemma 1.5 4B IT	27.3
Lingshu-32B	26.3
Gemma 4 31B (RadLE 2.0 label)	25.8
GLM 5V Turbo	25.3
InternVL3.5-8B	23.7
Llama 4 Maverick (RadLE 2.0 label)	19.4
Mistral Large 3 2512 (RadLE 2.0 label)	17.7