StateSight Benchmark Shows GPT-5.5 at 59.3% on Cube-Net Reasoning, Below 80.8% Human Baseline

The StateSight benchmark isolates latent spatial-state reconstruction through three procedural task families, each with 300 single-image prompts and deterministic oracle labels. OpenAI GPT-5.5 and Claude Sonnet 5 produced format-valid outputs yet recorded the listed accuracies. Human participants achieved 80.8%, 68.8%, and 64.3% mean accuracy on matched subsets. Visible-derivation traces documented systematic failures to recover hidden cube orientations and connectivity relations before reasoning began.

StateSight-Steps supplies 900 interleaved image-text sequences containing 3,600 intermediate visual states for step-level inspection. The gap between format compliance and spatial recovery indicates that current multimodal training objectives do not enforce explicit reconstruction of occluded geometry. Related work on visual spatial reasoning datasets shows similar dissociation between surface token prediction and internal state tracking.

Operationally, systems that score below 40% on tower-counting and connectivity tasks cannot be assumed to maintain consistent world models across partial observations. Integration into planning or manipulation pipelines will require auxiliary reconstruction losses or explicit 3D state supervision rather than reliance on end-to-end VQA accuracy.

Future releases of StateSight will add rotation-equivariant variants and dynamic occlusion sequences. Model developers can use the companion step dataset to train or evaluate intermediate visual state prediction before final answer generation.