Medical Information Response Audit (MIRA) - Completeness: leaderboard

Completeness loss (1-5) in LLM answers to medical questions, judged against the essential information each seed question requires. MIRA (2026) crosses 60 medically reviewed low-risk questions with Chinese or English, formal or colloquial, and high or low health-literacy phrasing; an LLM judge scores each answer on a 1-5 rubric where higher is worse. Lower is better.

Metric: Completeness loss (Q2), essential information missing against the per-seed checklist, LLM-judge rubric score (1-5), where higher is worse, averaged over the 4,320 MIRA prompts without mitigation: 60 medically reviewed low-risk seed questions crossed with language (Chinese, English), register and health-literacy signals, three prompt formats and three framings; lower is better. Source: arxiv.org. Saturation forecast: Around February 2027. 5 models tracked. Updated 7 Oct 2026.

Top models

#ModelScoreOverall rank
1DeepSeek V4 Pro2.27#196
2Qwen 3.6 Plus2.34#238
3GPT-5.42.37#135
4Claude Sonnet 4.62.4#173
5Llama 3.3 70B Instruct3.03#1385

No result here: #1 Claude Fable 5.2, #3 Gemini 4 Neon, #4 GPT-6.2 Astra.

Interactive version: theaggregate.ai/benchmark?slug=medical-information-response-audit-mira-completeness · How It Works · Data refreshed daily, snapshot 2026-11-21.