PhysiBench

A physics-grounded visual reasoning benchmark testing whether AI vision models can judge the physical plausibility of projectile motion.

Inspired by the labeling rigor of ImageNet and the spatial-intelligence focus of Fei-Fei Li's World Labs — PhysiBench targets a narrower, more measurable question: can current vision-language models tell when a trajectory violates real physics?

Dataset

• 300 labeled items, each a rendered PNG image of an object's position at 6 sampled moments in time, plus a JSON metadata record.
• 180 valid trajectories — real projectile motion, varying launch speed (5–25 m/s), angle (15–75°), and gravity (Earth/Moon/Mars). ~35% simulate real quadratic air resistance (drag coefficient Cd=0.47, the real drag coefficient of a sphere, integrated via small time-step simulation rather than a simplified formula).
• 120 invalid trajectories — 24 examples each of 5 distinct physics violations:
  • gravity_inversion — accelerates upward instead of falling
  • energy_injection — unexplained mid-flight speed increase
  • discontinuous_teleport — sudden position jump between frames
  • time_rewind — path partially reverses
  • impossible_landing — passes through the ground plane
• Splits: 210 train / 45 val / 45 test (70/15/15), stratified across categories.

Baseline Results

Evaluated on Gemini 3.6 Flash, 45-item test set. Full analysis in REPORT.md.

Overall accuracy: 86.7% (39/45)

Category	Accuracy
Valid trajectories	21/23 (91.3%)
gravity_inversion	0/4 (0.0%)
energy_injection	4/4 (100%)
discontinuous_teleport	6/6 (100%)
time_rewind	5/5 (100%)
impossible_landing	3/3 (100%)

Headline finding: Gemini 3.6 Flash perfectly detects violations of energy conservation, spatial continuity, causality, and ground-boundary constraints — but completely fails to detect inverted-gravity trajectories (0/4), consistently describing them as standard projectile motion despite using correct gravitational vocabulary in its own explanation. See REPORT.md for full error analysis.

Files

generator/
  physics.py          — ground-truth trajectory generator (clean + real drag physics)
  violations.py        — 5 labeled physics-violation functions
  render.py             — renders trajectories as PNG images
  generate_dataset.py   — orchestrates generation of all 300 items
eval/
  run_baseline.py       — evaluates a vision model (Claude) against the test split
data/
  images/                — 300 PNG files
  dataset.jsonl          — one JSON record per line: id, valid, violation_type, params, image_path, split

Loading the dataset

import pandas as pd
df = pd.read_json("data/dataset.jsonl", lines=True)

Running the baseline evaluation

pip install anthropic
export ANTHROPIC_API_KEY="your-key-here"
cd eval && python3 run_baseline.py

License

Dataset and code released under CC-BY 4.0. Free to use, share, and adapt with attribution.

Citation

If you use PhysiBench in your work, please cite this repository (see CITATION.cff).