FireWorldBench Read paper
A BENCHMARK FOR PHYSICAL WORLD INTELLIGENCE

FireWorldBenchBenchmarking Complex Physical World Intelligence through
Coupled-Field Fire Dynamics

Can multimodal models understand a world of interacting physical fields?

Qiang Chen1★†Hao Guo2★Huatai Zhu2★Tairan Huang2★Yichao Cao2

Hongyan Xu2Keke Huang2Haifeng Li2Yi Chen1‡Xiu Su2‡

1 Hong Kong University of Science and Technology2 Central South University

★ Core contributors   † Project lead   ‡ Corresponding authors

Paper Code Dataset Results

Code and dataset links are listed in the manuscript; public access is currently unavailable.

KEY TAKEAWAY

Recognizing a fire is only the beginning.
Understanding how physical fields interact remains a challenge for today’s models.

MULTIPLE VIEWS OF FIRE DYNAMICS.

From visible fire to invisible dynamics.

Paper · Figs. 1 & 5
Building fire photograph from the Real Fire Events panel in Figure 1 of the paper
01Observed eventReal-world evidence
Aligned three-dimensional reconstruction of the fire scene
023D reconstructionScene geometry
Temperature field from the aligned fire simulation
03TemperatureThermal evolution
Soot density field from the aligned fire simulation
04Soot densitySmoke transport

Real fire example: Fig. 1. Reconstruction and physical fields: Fig. 5.

520Fire worlds
47Scene archetypes
9,074Question–answer pairs
5Physical capabilities
7Environment families
01 / THE BENCHMARK

Physical intelligence,
under a different kind of test.

Beyond objects and appearances: reasoning about continuous fields, hidden mechanisms, and a world that changes over time.

FireWorldBench is a multimodal benchmark for complex physical-world intelligence through coupled-field fire dynamics. Across 520 fire worlds and 9,074 question–answer instances, it combines structured observations, physical-field visualizations, and 3D scene representations.

A dual-axis evaluation tests five physical capabilities across three stages of fire evolution, using matched choice and open-report interfaces. The results reveal persistent gaps in reading physical fields, connecting observations to causal mechanisms, and reasoning about interventions.

P1

Temporal
forecasting

What happens next?

Predict how fire and physical fields evolve from partial observations.
P2

Perception
& grounding

What is happening, and where?

Locate and interpret physical states within the scene.
P3

Cross-field
coupling

How do the fields interact?

Connect heat, smoke, airflow, and visibility as one system.
P4

Causal
attribution

Why does it happen?

Explain observed changes through their physical mechanisms.
P5

Counterfactual
intervention

What if conditions change?

Reason about interventions and their downstream effects.
Across three stagesT1 Localized onsetT2 Coupled propagationT3 Critical transition
Figure 1. From multimodal fire worlds to capability-resolved evaluation.
02 / EXPLORE THE FIRE WORLDS

One physical world.
Many ways to observe it.

Co-registered 3D scenes and physical fields reveal how a fire develops across seven environment families.

TRANSPORTATION AND INFRASTRUCTURE

Electric bus terminal charging-bay fire

Controlled simulation

Follow the buoyant plume, bay-scale confinement, and longitudinal smoke transport through a charging-bay fire.

Paper · Fig. 16
494

Controlled simulation worlds

Reproducible scenarios with aligned geometry, physical fields, and temporal observations.

26

Real-world-aligned event groups

Observed experimental fires anchor reconstructions, with latent fields completed through simulation.

03 / EVALUATION & FINDINGS

Seeing the fields
isn’t yet understanding them.

A complementary evaluation of structured physical records and multimodal visual evidence.

Click a heading to sort
Frontier model completion accuracy on controlled simulation, averaged over five physical capabilities
GPT-5.6-Luna55.34BEST42.9912.35
Qwen3.5-Plus-02-1547.2539.407.85
Gemini-2.5-Flash46.2044.83BEST1.37
Claude-Haiku-4.544.7839.515.27

Completion accuracy (%), averaged over P1–P5 and aggregating choice and open-report scores. Frontier-model excerpt from Table 1; best scores are highlighted within this selection. The S − I gap is in percentage points.

01

A visual evidence gap

All four frontier models score lower on rendered physical evidence than on structured records in the reported capability averages.

02

Answers need mechanisms

Evidence-F1, Mechanism Alignment, and Gold-Linked Support evaluate whether explanations connect predictions to physical evidence.

03

Adaptation helps

FireWorldGPT studies physics-aware prompting, chain-of-thought reasoning, and supervised adaptation to strengthen domain understanding.

FIREWORLDGPT

A path toward better
physical understanding.

Supervised adaptation improves completion accuracy across all four evaluated base models.

+12.0percentage points
largest accuracy improvement

Vanilla+ Supervised fine-tuning

Controlled simulations · Table 5 · Accuracy (%)
Accuracy gains do not imply improved confidence calibration.

04 / FROM WORLDS TO QUESTIONS

Grounded in physics.
Evaluated along two axes.

Designed scenarios and observed experiments flow through a shared, provenance-preserving pipeline.

  1. 01

    Construct the world

    Define geometry, materials, ignition, and event constraints.

  2. 02

    Resolve the physics

    FDS simulates coupled fields; Smokeview produces aligned 3D renderings.

  3. 03

    Ask, predict, explain

    Matched questions test five capabilities across three fire stages, scored without an LLM judge.

Figure 4. A shared physical state connects the representations.
BUILD ON THIS WORK

Citation

If you find FireWorldBench useful in your research, please cite the manuscript.

Citation for the arXiv preprint.

BibTeX
@misc{chen2026fireworldbench,
  title  = {FireWorldBench: Benchmarking Complex Physical
            World Intelligence through Coupled-Field
            Fire Dynamics},
  author = {Qiang Chen and Hao Guo and Huatai Zhu and
            Tairan Huang and Yichao Cao and Hongyan Xu and
            Keke Huang and Haifeng Li and Yi Chen and Xiu Su},
  year   = {2026},
  eprint = {2609.23064},
  archivePrefix = {arXiv},
  primaryClass  = {cs.AI},
  url    = {https://arxiv.org/abs/2609.23064}
}