Behavioural intelligence for evacuation

We model the human factors
disaster simulations miss.

LLM agents simulate how people think · an agent-based model simulates how the world changes.

scroll ↓
See it live

Agents evacuating on the real Noto road network.

Run the Noto simulation to populate the live map.
Built on real dataESSD building damageOpenStreetMapGSI / JSGS inundationJMA warningsNanao survey · Hasegawa 2025Discovered: neighbours fleeing → +50% evacuation (causal, robust)
Motivation

Three assumptions existing models get wrong.

01
Everyone evacuates immediately?

Real data: departure delays of 0–30 minutes, varying significantly by demographic group.

02
People follow the shortest path?

Under uncertainty, people follow crowds or head to familiar routes rather than optimal ones.

03
Decisions are a black box?

Rule-based ABMs cannot explain individual behaviour; planners cannot interpret model outputs.

Our idea

A Disaster World Model driven by LLM agents.

01
Environment Engine

Converts raw data into a structured world state, updated each timestep.

02
LLM Agent Population

Reads the world state, reasons in natural language, and acts. Movement updates the state.

03
Calibration & Validation

Matches simulated population patterns against real human-mobility data.

Grounded in real data

The 2024 Noto Peninsula event.

Three layers — Data → Environment → Agent — built on the real earthquake-tsunami: building damage, OSM roads, GSI/JSGS inundation, designated shelters, and the JMA warning timeline.

  1. 1
    Data Layer

    Building damage · road network · tsunami inundation · shelters · warning timeline · run-up survey.

  2. 2
    Environment Engine

    Converts the raw layers into a structured world state at each timestep — passability, flooding, capacity, warning level.

  3. 3
    LLM Agent Population

    Survey-initialized personas Observe → Think → Act, emitting natural-language reasoning for every decision.

Noto Peninsula, Ishikawa · 37.4°N 137.2°E
Sandbox

Run the whole loop, in one board.

Drive the pipeline and inspect what it discovered — the hidden variables, their causal effects, agent reasoning, and the calibration — live.

Pipeline
0
Agent decisions
0
Hidden variables
0
Causal effects
Run the Noto simulation to populate the live map.