MindTopo: Multimodal Benchmark for Topological Reasoning in Foundation Models

MLL-Lab created MindTopo, a benchmark dataset containing 8,910 procedurally generated examples across 13 environments and 5 categories. It probes whether foundation models reason about topological concepts like connectivity and knottedness rather than superficial visual cues. The dataset was last updated on May 7, 2026.

Use Cases

• Benchmarking model performance on topological reasoning tasks based on the described categories of connectivity, enclosure, knottedness, ordering, and separation.
• Analyzing whether models rely on superficial visual cues versus structural understanding based on the procedurally generated examples.
• Developing new evaluation metrics for cognitive science and neuroscience applications based on the multimodal benchmark design.

Strengths

• Dataset contains 8,910 examples, providing a substantial test set.
• Examples are procedurally generated across 13 environments, likely ensuring controlled variation.
• Benchmark targets 5 specific topological categories, providing a structured evaluation framework.

Limitations

• Column-level documentation is absent; field semantics must be inferred after download.
• Row count is unknown, which may limit suitability assessment.
• Description metadata is limited; actual data quality requires manual inspection after download.

Provenance

Source

MLL-Lab

Collection Method

Procedurally generated examples.

Freshness

Last updated 2026-05-07 10:58:33; freshness should be verified.