← All benchmarks

MA-Bench

Professional Work · 2026-10-13

Scale AI and Data Machines' benchmark of mission analysis, the first step of military operational planning (US joint doctrine, JP 5-0). For each of 10 expert-built scenarios the model reads a synthetic staff packet of 10 to 12 documents (35,649 to 242,658 characters) in one prompt and writes a mission analysis brief, cut at 3,000 words before grading. The judge, GPT-5.6 (served as gpt-5.6-sol) at high reasoning effort with a frozen prompt and no access to the packet, grades each brief in one call against the scenario's expert-written pass/fail rubric, 1,257 items in all (74 to 170 per scenario); a positive item is correct when the brief makes the required determination and a negative item, which names a wrong move, when the brief avoids it. The score is the paper's Correct, the percentage of rubric items graded correct, averaged over a model's briefs per scenario (three, or five for three models) and then over the 10 scenarios; the paper's headline severity-weighted score is not imported.

Top models (higher is better)