DORA - Multimodal Report Synthesis: leaderboard

DORA evaluates model agents answering disaster-response questions with a 108-tool geospatial library under a ReAct loop. This series reports final-answer accuracy for multimodal report synthesis; the five tasks cover 515 expert-authored cases across 45 disasters. Only the thirteen agent runs enter this panel; expert gold trajectories and tool-free VLM runs use different protocols. Higher is better.

Metric: Final-answer accuracy (%; ReAct agent with tools). Source: github.com. Saturation forecast: Around April 2027. 13 models tracked. Updated 2 Oct 2026.

Top models

#ModelScoreOverall rank
1Qwen 3.5 397B A17B49.17#255
2Claude Sonnet 4.645.53#173
3Gemini 3 Flash45.31#174
4Grok 4.1 Fast44.52#429
5MiMo-V2-Pro44.26#329
6Gemma 4 31B42.03#385
7Step 3.5 Flash41.58#520
8DeepSeek V3.241.01#417
9MiniMax-M2.740.62#533
10GPT-5.429.11#135
11GPT-OSS-120B28.84#707
12GPT-5.4 Nano27.37#645
13Qwen 3.5 35B A3B21.3#490

No result here: #1 Claude Fable 5.2, #2 GPT-6.2 Astra, #3 Gemini 4 Carbon.

Interactive version: theaggregate.ai/benchmark?slug=dora-multimodal-report-synthesis · How It Works · Data refreshed daily, snapshot 2026-11-17.