Spatial Reasoning Eval

spatial-reasoning-eval

Geometrically-Constrained Agent for Spatial Reasoning — Chen et al. (2025) (arXiv:2511.22659, 2025)

What this evaluates

This evaluation probes a model's ability to perform complex spatial reasoning and perspective-taking across single and multiple images. It specifically tests whether the model can correctly establish geometric reference frames, handle multi-step transformations, and generalize across different spatial logic tasks without relying on dataset-specific biases.

Datasets

• MMSI-Bench — total ?; splits: test (-1)
• MindCube-tiny — total ?; splits: test (-1)
• OmniSpatial — total ?; splits: test (-1)
• SPBench — total ?; splits: test (-1)
• CV-Bench — total ?; splits: test (-1)

Metrics

• overall accuracy (%) (primary) — range: percent
  • Percentage of correctly predicted answers out of the total number of instances in the benchmark.

Input / output format

Input: Single or multiple images paired with spatial reasoning questions/prompts (e.g., 4-choice multiple choice questions for MMSI-Bench).

Output: A predicted answer choice or generated response corresponding to the spatial reasoning query.

Scoring recipe

def compute_accuracy(predictions, gold_labels):
    correct = sum(1 for p, g in zip(predictions, gold_labels) if p == g)
    return (correct / len(gold_labels)) * 100

Common pitfalls

• Models often default to flawed semantic shortcuts or lossy spatial imagination instead of establishing formal geometric constraints.
• Training-based methods frequently exhibit strong in-domain performance but fail to generalize to out-of-domain benchmarks due to dataset-specific biases.
• Unconstrained tool integration without explicit task constraints leads to unreliable planning and incorrect geometric computations.

Evidence (verbatim from paper)

For all benchmarks, we report both overall accuracy (%) and subcategory accuracy (%).

Citation

@misc{chen2025geometricallyconstrained,
  title={Geometrically-Constrained Agent for Spatial Reasoning},
  author={Chen et al. (2025)},
  year={2025},
  note={arXiv:2511.22659}
}

• arXiv: 2511.22659