Disaster Mitigation Bench

The Disaster Mitigation Bench benchmarks language models as drone navigators and reconnaissance advisors in simulated 3D disaster environments. It covers 15 scenarios across 6 disaster types (forest fire, peat fire, earthquake, flood, flash flood, landslide), built from real data: 391,641 satellite fire hotspots, real terrain elevation, and Sentinel-2 imagery, with deterministic seeded worlds so every run is reproducible. Two roles are scored. The navigator flies a drone to find victims under tight frame and scan budgets. The recon advisor writes rescue plans, graded against hidden truth by waypoint compliance plus human review. The bench runs on two sim backends, Gazebo with real flight-control software and Unreal Engine. Headline results: on the base pack the frontier model scores 0.755 while scripted coverage beats the 7B model, but on the harder worlds every model finds zero victims in 18 FIRMS runs, and quake plus slide yield exactly 1 find in 32 runs. All findings, including null results and failure modes, ship with receipts in the repo.