RadLE 2.0 Reliability Index (RadLE-R)

Multimodal

RadLE-R measures diagnostic accuracy among answers submitted with autonomous-level confidence (Likert 3 or 4). Its denominator is each participant's high-confidence answers and varies by participant; it is not the full 200-case exam denominator used by RadLE-A. It tests whether the answers a system presents most confidently can be trusted.

Top models (higher is better)

Model	Score
Human Expert Baseline (RadLE 2.0)	55.0
Claude Fable 5	54.7
Muse Spark 1.1	40.2
Gemini 3.1 Pro Preview	32.5
GPT-5.6 Sol Pro	31.9
Qwen3.7-Plus (source unspecified)	18.7
GLM 5V Turbo	17.2
Grok 4.5	15.5
Gemma 4 31B (RadLE 2.0 label)	9.4
Llama 4 Maverick (RadLE 2.0 label)	7.9
MiniMax M3	7.3
OctoMed-7B	3.5
Lingshu-32B	3.5
Nemotron 3 Nano Omni 30B A3B	0.9
MedGemma 1.5 4B IT	0.7
InternVL3.5-8B	0.6
Mistral Large 3 2512 (RadLE 2.0 label)	0.5