GEB-Bench

Abstract Structures Told in Many Voices

Tong Zhang1, Zhiyuan Shi, Yun Peng1, Tao Xie2,∗

1Fudan University    2Peking University    ∗Corresponding author

Paper arXiv Code GitHub Cite BibTeX

Analogy requires more than naming a familiar relation. It requires carrying that relation into a new situation with different objects and content. GEB-Bench studies this ability by expressing the same abstract motif as a scene, a story, a mathematical statement, or a programmatic skeleton.

GEB-Bench overview showing one motif expressed in four voices
One relational motif can be realized through very different content. Identification names the motif in one voice, while matching connects two voices by their shared motif.

questionCan a model carry a relation across content?

Equivalent representations preserve the same problem in another notation. Analogy is harder because the objects and subject matter may change while only the relation remains. GEB-Bench makes that distinction testable.

Scene
The relation appears through spatial composition and visible organization.
Story
The relation appears through order, repetition, and the form of the telling.
Mathematics
The relation appears in a mathematical statement without relying on shared surface content.
Skeleton
The relation appears as a compact programmatic construction.

tasksIdentification and matching

Identification

A model receives one item and selects its motif from the full motif menu. This tests whether the structure is recognizable in each voice.

Cross-voice matching

A model receives a source item and selects the candidate in another voice that expresses the same motif. Scene-to-story, scene-to-mathematics, and story-to-mathematics pairs test whether a recognized relation can guide comparison across different content.

The construction pipeline starts from the relation specified for each motif, generates every voice from that relation, checks the order and repetition of story units, and assembles linked identification and matching questions. The linked design lets us compare a model's component judgments with its final match.

findingsIdentification does not guarantee analogy

Models can identify the motif in a scene and in each candidate story yet still choose the wrong match. Recognizing a relation and carrying it across different content are distinct abilities.

libraryRelations rather than objects

The motif library spans self-reference, recursion, cycles, symmetry, tiling, graph structure, and counting arguments. Each item is generated from a motif-specific relation rather than from a shared object template. A recursive river delta and a recursive story may share no objects or vocabulary, yet they can still instantiate the same organization.

Recursion skeleton
recursion skeleton
River delta scene
river delta scene
Lightning scene
lightning scene
Self-reference skeleton
self-reference skeleton
Self-reference scene
self-reference scene
Cycle skeleton
cycle skeleton

resourcesPaper, dataset, and code

The public repository contains the motif library, generated voices, evaluation questions, and analysis code used by the benchmark.

github.com/tongmd/GEB-benchmark

citationBibTeX

@misc{zhang2026gebbench,
  title         = {GEB-Bench: Abstract Structures Told in Many Voices},
  author        = {Zhang, Tong and Shi, Zhiyuan and Peng, Yun and Xie, Tao},
  year          = {2026},
  eprint        = {2608.04111},
  archivePrefix = {arXiv},
  primaryClass  = {cs.CV},
  url           = {https://arxiv.org/abs/2608.04111}
}