Datasets:
The dataset viewer is not available because its heuristics could not detect any supported data files. You can try uploading some data files, or configuring the data files location manually.
DominoBench: Evaluating Multi-Step Physical Reasoning from Still Images
Paper: arXiv:2611.20385
A ball is pushed. It sets off a row of dominoes, a small ball rolls on into a second row, and a last ball drops into the bucket of a seesaw whose spikes hang over a person. Who dies? Every DominoBench item is one still image of such a 2D rigid-body chain reaction, and the answer comes from running the scene in a physics simulator.
Items come in twins: two scenes identical except for one quantity (a drop height, a gap, a counterweight, a notch position), moved just across the threshold where the outcome flips. A model that tells a plausible chain-reaction story without checking that quantity gives the same answer to both twins and gets one of them wrong.
A minimal pair (pair 4): the counterweight on the seesaw is one grid step wider in (b). In (a) the ball in the bucket outweighs it and the spikes come down on C; in (b) the seesaw does not tip.
Each item also carries a hint ladder of four prompts, from no help to the written-out deciding calculation:
| Rung | What the prompt adds |
|---|---|
| H0 | nothing: "Who dies if A pushes the ball?" |
| H1 | a generic nudge to check the quantitative details in the picture |
| H2 | a pointer to the deciding step |
| H3 | the deciding calculation and the quantities it needs, including quantities a still image cannot show |
What is in this repository
| Path | Contents |
|---|---|
data/dominobench1k/ |
DominoBench-1K: 854 items (427 twin pairs), seven deciding mechanisms |
data/dominobench1k/dominobench124_pairs.json |
DominoBench-124: a 62-pair subset (124 items), the main ranking set |
data/dominobench1k/dominoscore_weights.json |
the frozen DominoScore weights |
data/m1/ |
DominoBench-38: 38 items (19 twin pairs), six mechanisms, a small quick-to-run test set |
data/excluded_pairs.json |
the generated pairs that were removed, and why (see below) |
dominobench/ |
the generator: parts, scene composition, simulation, labelling, rendering, guided search |
eval/ |
the evaluation harness: prompts, model backends, answer parser, metrics, runner |
Each item has three files:
<item_id>.png: the still image (grid drawn at 1 unit = 1 square)<item_id>.json: question, ground-truth answer (who dies / survives), the twin's id, the deciding mechanism and parameter, the four hint prompts, a structured scene description for text mode, contacts, chain length, the naive chain-reaction answer, and the scene program that regenerates the scene<item_id>_trajectory.npz: the simulated trajectory
index.json lists every item with its twin, seed, answer, deciding mechanism and the flipped parameter value.
The split field inside item files is an internal generator label; use the folder to tell the benchmarks apart.
The seven mechanisms are dominoes, jump, ledge, pendulum, pit, recess and seesaw. Every item passed a robustness filter: its outcome is unchanged when the timestep is halved and solver iterations are doubled, and under friction x0.9 / x1.1 and elasticity +0.05.
Removed pairs: every death is visible in the picture
A person dies when a moving body enters their collision box, which reaches slightly beyond the drawn figure near its
corners. We checked every labelled death against the drawn figure and removed the pairs in which a death comes from a
body that enters the box without touching the figure, so that a model that reads the picture exactly is never marked
wrong: 73 of the 500 generated pairs of DominoBench-1K (45 of them at ledges) and 6 of the 25 pairs of the small set.
The remaining pair numbers are kept, so numbering has gaps. data/excluded_pairs.json lists the removed pairs.
The DominoScore weights were computed after this filter.
Scoring
Models are scored on three numbers per item set: exact (every person's fate right), deciding person (the fate
of the person whose outcome flips between the twins) and both twins (both items of a pair exactly right).
DominoScore (0 to 100) is a chance-corrected, mechanism-weighted score of deciding-person accuracy: each
mechanism's skill above chance is clipped at zero and weighted by the frozen weights in dominoscore_weights.json, so
harder mechanisms count more and later models are scored on the same scale. See eval/metrics.py.
Running the evaluation
pip install -r requirements.txt
# DominoBench-1K, image mode, unaided and with the calculation, one backend:
python eval/run_eval.py --backend claude:sonnet --dataset 1k --modes image --rungs H0 H3 --pairs 1-500
# DominoBench-124:
python eval/run_eval.py --backend claude:sonnet --dataset 1k --modes image --rungs H0 H3 \
--pairs @data/dominobench1k/dominobench124_pairs.json
# DominoBench-38, full ladder, image and text mode:
python eval/run_eval.py --backend claude:sonnet --dataset m1 --modes image text --rungs H0 H1 H2 H3 --pairs 1-25
Pair ranges select the pairs that exist, so --pairs 1-500 runs the 427 pairs of DominoBench-1K.
Backends (eval/backends.py): claude:<model> (Claude Code CLI, tools disabled), codex:<model> (Codex CLI,
read-only sandbox), antigravity:<model> (Antigravity CLI, with the evaluation-only deny rules in
eval/agy_home/ and the audit in scripts/agy_tool_audit.py), gemini:<model>, xai:<model>, openrouter:<model>,
and baseline:{oracle,all_survive,chain,naive}. API keys are read from ~/.config/dominobench/keys.env. Open-weight
models run through vLLM with eval/colab_vllm_runner.py and are imported with eval/colab_io.py. Answers are written
to results/ and the run is resumable.
Drawing new items
The generator draws new scenes from held-out seeds, so fresh items can replace the released ones whenever
contamination becomes a concern. Simulations are deterministic (fixed timestep, seeded) and use pymunk.
Citation
@misc{ferrara2026dominobench,
title = {DominoBench: Evaluating Multi-Step Physical Reasoning from Still Images},
author = {Ferrara, Emilio},
year = {2026},
eprint = {2611.20385},
archivePrefix = {arXiv},
primaryClass = {cs.CV},
url = {https://arxiv.org/abs/2611.20385}
}
License
MIT, for the data and the code.
- Downloads last month
- 32

