bmrayan/perception-not-reasoning

Perception, Not Reasoning, Limits Visual Theory of Mind

Accepted to NeurIPS (E&D) 2026. Oral presentation at the BabyVLM Workshop, NeurIPS 2026.

Mohamed Rayan Barhdadi, Syed Talal Wasim, Juergen Gall, Erchin Serpedin, and Hasan Kurban.

Code | Website

This release contains the paper's original abstract 2D stimuli, annotations, and saved evaluation results. It includes 1,000 probes across 20 templates, with 50 variants per template. All probes belong to the test split.

Dataset

Stage	Task	Templates	Probes
1	Object tracking	6	300
2	Goal attribution	4	200
3	False belief	10	500

Load

from datasets import load_dataset

data = load_dataset("bmrayan/perception-not-reasoning", split="test")
print(data[0]["question"], data[0]["answer"])

Files

Path	Contents
images/	PNG sequences and ImageFolder annotations
probes.json	Prompts, answers, seeds, scene metadata, and scaffolds
metadata.csv	Flat annotations
images.zip	Image archive referenced by Croissant
croissant.json	Dataset schema, download URLs, and responsible-use metadata
checksums.json	SHA-256 file hashes
results/reference/	Saved paper responses and scores
results/followup/	Additional control results
results/provenance.json	Original result-file hashes

The default loader reads the image dataset. Result bundles are separate files.

Croissant metadata:

• Abstract probes and images.
• MuMA-ToM evaluation records.
• MindPower evaluation scores.

The naturalistic metadata describes saved evaluation outcomes, with source-file and row references. It does not redistribute the original videos or simulator assets. Obtain the original benchmarks from MuMA-ToM and MindPower.

Fields and construction

probe_id, stage, template, and panel_count identify each task. question, answer, and answer_kind specify the prompt and grading key. scaffold_percept, scaffold_full, and text_narration contain the text conditions.

The deterministic generator is evaluation/abstract2d/evaluate.py in the code repository. Variants are indexed 0-49. Seeds follow 10000 * (template_index + 1) + variant_index. Questions, answers, and descriptions derive from the same scene state.

The 1,000 images match the original experiment files byte for byte. Prompts and answer keys match all 35,000 saved seven-model phase-1 rows. Packaging changes image paths only.

Use and limitations

The dataset supports comparisons of input conditions in controlled visual theory-of-mind tasks. It is synthetic, English-only, and template-based. Repeated layouts and answer conventions can support shortcuts; template variants are not independent concepts.

The permanence and color-change Stage 1 prompts explicitly include their expected answers. Those items do not independently measure visual understanding. Full scaffolds contain more information than perceptual scaffolds. Report conditions, model panels, and scoring rules separately.

Scenes contain fictional geometric characters. No participant data or personal information were collected. Simplified scenes and English-only prompts limit population and cultural coverage. The scores do not establish general social competence and are not validated for clinical assessment or decisions about people.

License

Original abstract images and annotations: CC BY 4.0, as specified in LICENSE. Code: MIT. Naturalistic source media are not included. MuMA-ToM benchmark text retains MIT terms; MindPower reference text retains Apache-2.0 terms. See third-party attribution and license notices. Simulator software and assets are not included.

Citation

@inproceedings{barhdadi2026perception,
  title     = {Perception, Not Reasoning, Limits Visual Theory of Mind},
  author    = {Barhdadi, Mohamed Rayan and Wasim, Syed Talal and Gall, Juergen
               and Serpedin, Erchin and Kurban, Hasan},
  booktitle = {Advances in Neural Information Processing Systems},
  year      = {2026}
}