DisasterBench - RAP: leaderboard

Reasoning via Planning, a Monte Carlo tree search guided by usefulness and self-consistency rewards, on DisasterBench (2026), which tests whether an LLM can compose executable disaster-response workflows from 26 specialized tools under typed interface constraints. Exact-match accuracy over the planning tasks (every step must match in tool, parameter bindings and dependencies), percent; higher is better.

Metric: Exact-match accuracy (%) over the 233 DisasterBench planning tasks (35 single-tool, 166 chain and 32 branching workflows over 26 disaster-response tools): a plan counts only when every step matches the ground truth in tool identity, parameter bindings and inter-step dependencies; planning method: Reasoning via Planning, Monte Carlo tree search guided by usefulness and self-consistency rewards (temperature 0.5); higher is better. Source: arxiv.org. Saturation forecast: Around February 2027. 13 models tracked. Updated 7 Oct 2026.

Top models

#ModelScoreOverall rank
1Gemini 3.1 Pro (Preview)72.1#99
2GPT-5.465.67#135
3Qwen 3 Max63.09#415
4Qwen 3.5 27B58.8#375
5Gemma 4 31B50.64#382
6Qwen 3.5 9B42.92#828
7DeepSeek R135.19#482
8DeepSeek V3.234.33#419
9Llama 3.3 70B28.76#1514
10Llama 3.1 8B28.33#5196
11Ministral 3 14B18.88#1870
12Ministral 8B (DisasterBench checkpoint unspecified)12.88
13Ministral 3B (DisasterBench checkpoint unspecified)3.86

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

Interactive version: theaggregate.ai/benchmark?slug=disasterbench-rap · How It Works · Data refreshed daily, snapshot 2026-11-21.