Obshazard-bench: leaderboard
Obshazard-bench (National University of Defense Technology, Shanghai AI Laboratory and Shanghai Jiao Tong University, 2026) evaluates multimodal foundation models on disaster intelligence from raw satellite sounding streams combined with ground-station observations, historical disaster records and socio-economic indicators, over 4,202 visual question-answering samples from 127 extreme events in eight disaster categories. Answers get a task-aware soft score from 0 to 1 (categorical answers by semantic or lexical consistency, numeric and interval answers by distance to the truth), averaged with equal weight over timesteps, subtasks, lifecycle stages and disaster categories and shown in percent; higher is better. Models answer zero-shot with uniform prompts.
Metric: Normalized task score (%; x100 of the 0-1 task-aware soft score following the Earth AI protocol: categorical answers by semantic or lexical consistency, numeric and interval answers by distance to the ground truth; equal-weight aggregation across timesteps, subtasks, lifecycle stages and disaster categories; zero-shot, uniform prompts; raw satellite sounding streams with ground-station, disaster-record and socio-economic inputs; average of the three lifecycle stages over 4,202 VQA samples from 127 historical extreme events in 8 disaster categories). Source: arxiv.org. Saturation forecast: Around April 2028. 4 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | GPT-5.5 | 28.33 | #48 |
| 2 | Kimi K2.6 | 25.06 | #173 |
| 3 | Claude Opus 4.8 | 24.24 | #56 |
| 4 | Qwen 3.5 397B A17B | 19.22 | #242 |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=obshazard-bench · How It Works · Data refreshed daily, snapshot 2026-11-12.