Multi-modal event analysis and social simulation✨

A lightweight system for turning event materials into structured graphs, agent populations, and inspectable social simulations.

Project Site · Architecture · User Guide · Examples

MicroWorld starts from raw event material: 📄 documents, 🖼️ images, 🎥 videos, and 🕸️ graph-ready context. It builds an event graph, derives platform-facing agent profiles, runs a multi-agent discussion process, and keeps the intermediate artifacts available for inspection after the run!

🎯 Why MicroWorld? It is built for cases where the final report is not enough on its own. The project keeps the graph 📊, prompts 💬, simulation inputs 🎬, action traces 👣, memory states 🧠, and report outputs 📑 tied to the same run.

MicroWorld is organized around four key stages:

1. 📥 Ingestion and graph build: Convert event materials into ontology, entities, and relations.
2. ⚙️ Simulation preparation: Derive topic keywords, cluster topology, and generate platform profiles.
3. 🏃‍♂️ Runtime execution: Run the topology-aware simulation with directional influence and lightweight memory.
4. 📈 Reporting and inspection: Collect logs, traces, configs, and reports from the same run.

Multi-modal Event Ingestion Handles text, image, and video inputs in one unified workflow.

Two Topology Clustering Modes Supports both threshold-based and LLM-keyword-driven clustering.

PPR-guided Directional Influence Models activation and information flow with topology-aware influence.

Lightweight Memory Preserves useful state incrementally without full-history replay.

Inspectable Outputs Keeps graph artifacts, traces, configs, and reports available for inspection.

💡 Efficiency Boost: Removes redundant updates and cuts token usage sharply as workloads grow. Keeps the simulation trajectory closer to the reference trend than the baseline run! 📉

Removes redundant updates and cuts token usage sharply as workloads grow.	Keeps the simulation trajectory closer to the reference trend than the baseline run.

The public example uses the LK-99 room-temperature-superconductor news cycle. It is a good fit for the project because it contains:

• 🧩 Mixed evidence types, including long-form text and videos.
• 🎢 A clear shift from early excitement to later scrutiny.
• 🗣️ Visible changes in narrative focus, participants, and discussion structure.

git clone https://github.com/d2i-cuhksz/MicroWorld.git
cd MicroWorld

Required:

• 🐍 Python 3.11+
• ⚡ uv

Recommended for video inputs:

• 🎞️ ffmpeg
• 🎞️ ffprobe

(If ffmpeg and ffprobe are not on your PATH, set MULTIMODAL_FFMPEG_PATH and MULTIMODAL_FFPROBE_PATH in .env)

cp .env.example .env

Set at least:

LLM_API_KEY=your_key
ZEP_API_KEY=your_key

Common defaults:

LLM_BASE_URL=https://dashscope.aliyuncs.com/compatible-mode/v1
LLM_MODEL_NAME=qwen-plus
MULTIMODAL_AUDIO_API_KEY=
MULTIMODAL_AUDIO_BASE_URL=

uv sync

uv run microworld-api

(The backend uses FLASK_HOST, FLASK_PORT, and FLASK_DEBUG from the environment. The default backend port is 5001.)

cp configs/full_run/full_run.template.json /tmp/microworld-run.json

Minimal example:

{
  "project_name": "My MicroWorld Run",
  "graph_name": "My MicroWorld Graph",
  "simulation_requirement": "Build entities, relations, and a two-platform social simulation from the input materials.",
  "files": [
    "/abs/path/to/event.md",
    "/abs/path/to/video.mp4"
  ],
  "pipeline": {
    "chunk_size": 500,
    "chunk_overlap": 50,
    "batch_size": 3
  },
  "simulation": {
    "enable_twitter": true,
    "enable_reddit": true
  },
  "report": {
    "generate": false
  }
}

(You can also leave files empty and provide files_from, with one local path per line.)

uv run microworld-full-run \
  --config /abs/path/to/microworld-run.json

To avoid the interactive clustering choice:

uv run microworld-full-run \
  --config /abs/path/to/microworld-run.json \
  --cluster-method threshold

Generated data is written under: 📁

• data/generated/
• output/simulations/
• output/reports/
• runs/

A typical run exposes artifacts from several stages:

• 🏗️ Project build: Extracted text, parsed multimodal content, source manifests, ontology output.
• 🎭 Simulation preparation: Entity prompts, graph snapshots, social relation graph, simulation config.
• 🎬 Runtime execution: Platform profiles, action logs, memory states, topology traces.
• 📊 Reporting: Report outline, full report, agent logs, console logs.

MicroWorld/
├── 📄 pyproject.toml
├── 📁 src/
│   └── 📁 microworld/
│       ├── 🔌 api/
│       ├── 📱 application/
│       ├── ⌨️ cli/
│       ├── ⚙️ config/
│       ├── 🧩 domain/
│       ├── 🕸️ graph/
│       ├── 📥 ingestion/
│       ├── 🏗️ infrastructure/
│       ├── 🧠 memory/
│       ├── 📊 reporting/
│       ├── 🏃 simulation/
│       ├── 💾 storage/
│       ├── 📡 telemetry/
│       └── 🛠️ tools/
├── 📁 configs/
│   ├── 🏃 full_run/
│   └── 🎬 simulation/
├── 📁 data/
│   └── 🏗️ generated/
├── 📁 docs/
└── 📁 tests/

Want to contribute or run tests?

uv sync --group dev
uv run pytest

MicroWorld is built on top of prior open-source efforts. We would like to thank MiroFish and OASIS for making their work publicly available. Our project extends and adapts ideas and implementation foundations from these repositories for a lightweight, topology-aware, multi-modal social simulation workflow.

If this repository is useful for your work, please cite our project.

@article{xu2026topology,
  title={Topology-Aware LLM-Driven Social Simulation: A Unified Framework for Efficient and Realistic Agent Dynamics},
  author={Xu, Yuwei and Zhang, Shulun and Zhou, Yingli and Zeng, Shipei and Lakshmanan, Laks VS and Ma, Chenhao},
  journal={arXiv preprint arXiv:2604.18011},
  year={2026}
}

MicroWorld is released under the GNU Affero General Public License v3.0.