Source Data
Manuscript: Vision-language models lag human performance on physical dynamics and intent reasoning
Tracking number: NCOMMS-26-036064C

Getting started
INDEX.txt maps each main figure, main table, Supplementary Figure 1–7 and Supplementary Table 1–16 to its supporting files. TXT tables use UTF-8 tab-separated values; CSV files use comma-separated values. JSON files contain structured analysis results. Files ending in _reported.txt reproduce the displayed, rounded table entries; the corresponding numeric tables and Analysis/ files provide the available counts and full-precision statistics.

Data and units
Model-comparison files contain the aggregate accuracies reported in the manuscript. Figure_2c.txt provides the six category accuracies for the eight plotted models, corresponding to Table_1_reported_model_scores.csv. Figure_4_human_counts.csv provides the number of correct answers, the denominator of 8000 questions and the resulting accuracy for each of eleven participants. Figure_3b_3d_counts.txt supplies the pipeline stage totals shown in the figure.

Accuracies are percentages unless a column explicitly indicates proportions. Gains in pp are percentage-point differences calculated before rounding. The SFT standard deviations describe two training seeds. The base condition has one evaluation and no seed standard deviation. Supplementary Table 7 uses descriptive Kish-adjusted Wilson intervals; Supplementary Table 8 intervals resample source videos. Supplementary Tables 2–3 describe the observed panel of models. NA/null indicates unavailable or inapplicable values; zero represents an observed zero. In Figure 5b, closed-model parameter counts are unavailable (NA); x=100 is only a display coordinate, not a measured or estimated parameter count.

Model labels
GPT5 and GPT-5 refer to the same model. Qwen-3-VL-Plus is the display label for qwen3-vl-plus. Figure 7 includes aggregate inputs, standardized values and clustering linkage data.

Benchmark categories
Figure_3c_category_assignments.txt supplies all 8000 category assignments and identifies the field containing each label. Five labels occur outside the C field: indices 2452, 2606, 3016 and 5160 belong to category 2; index 5479 belongs to category 4. Category counts are 1086, 2491, 662, 1215, 211 and 2335.

Coverage analyses
Only definite coding assignments contribute to Supplementary Tables 13–16. The union of the two 200-question sets contains 92 observed fine-grained types. Supporting analysis files retain coding status and tentative assignments. Coding is model-assisted and based on question text. Supplementary Table 16 includes the concise translations reported in the paper together with the original questions.

File locations and integrity
SOURCE_PROVENANCE.txt identifies supporting files within this package or the public Zenodo code archive (version v1.0.0, https://doi.org/10.5281/zenodo.23877838). Paths marked Source Data are relative to this directory; paths marked Zenodo are relative to the top-level directory of the archived repository. SHA-256 values identify the referenced file contents. MANIFEST.txt records the size and SHA-256 of each packaged file except itself. Supplementary Table 1 is a qualitative literature comparison.
