LayoutBench

What stands out

Text justification: 215/272 (79.0%); Grid gutter comparison: 163/204 (79.9%); Horizontal versus vertical inset: 164/204 (80.4%). These have the lowest pooled accuracy across this model roster. Pooling describes these responses; the graph and table show each model separately. Families have two to six answer choices, so chance levels differ; stimulus difficulty also limits comparisons between families.

Immediate flow, Text columns, Grid columns, Grid rows, Wrapped row count, Text block placement, Relative panel width, Top-level hierarchy, Nested section flow reached 100% across every compared configuration. A ceiling on these controlled images shows where this corpus stops distinguishing models. It does not establish general layout understanding.

Right-aligned text mistaken for centering

Right-aligned paragraphs were recognized in 42/68 responses; 16/68 instead called them centered. Claude Fable 5.1 and Claude Opus 5 and Gemini 3.8 Flash and GPT-6 Astra and Claude Opus 5.5 and GPT-6 Sol answered all sixteen text-justification questions correctly. The difficult right-aligned example below shows the exact text and all 17 answers.

Equal spaces draw more errors

Equal spacing produced more errors than unequal spacing in these questions. The table pools the same 17 completed configurations and keeps the two tasks separate.

GPT-5.6 Sol and GPT-6 Astra and GPT-6 Luna and GPT-6 Sol answered all twelve grid-gutter and all twelve inset questions correctly. These descriptive differences do not establish why the errors occurred.

A page can run vertically and a row horizontally

Layout changes with the level of the hierarchy being inspected. A page can stack sections vertically while the children of one section form a horizontal row or a wrapped collection. These three tasks name the level to inspect so a correct global description cannot substitute for the local answer.

Across 408 paired outcomes (17 models × 24 shared images), both answers were correct in 408/408; 0 had the outer answer correct and the inner answer wrong. Every compared configuration recognized both levels of flow in these outlined synthetic scenes.

Alignment within a wrapped section produced a different result: nested final-row alignment was correct in 398/408 answers. Claude Haiku 4.5 scored 17/24; Pareto scored 21/24; the other 15 configurations answered all 24 correctly. The corpus includes a left-aligned wrapped collection in the second vertically stacked section, so a local alignment question can differ from the page's overall flow.

In the nested example below, Claude Haiku 4.5 recognizes the vertically stacked page and the wrapped children inside the target section, but calls its left-aligned final row centered. The answers below show how the other configurations handled that same image.

Top-level hierarchy and nested section flow ask two questions about each of the same 24 images. Comparing their paired answers shows where a model recognizes the outer structure but misses its children. Flat and nested stimuli still do not isolate a causal effect of nesting.

Inspect the questions and mistakes

Each example shows a selected question from its family. The expandable answers show successes and errors on the same image. Models receive the image and question, including its answer choices, without the ground truth.