Vision-language models may solve abstract visual comparisons through two competing internal processes: an early one that follows visible objects and a later one that represents relations between them.
That is the central finding of a preprint by Minegishi, Furuta, Kojima and colleagues. The team adapted a Relational Match-to-Sample task from developmental and comparative psychology, then tested both closed frontier systems and open models on controlled images.
Why it matters
For a researcher evaluating visual reasoning, a correct answer alone cannot reveal whether a model matched the underlying rule or merely recognized similar shapes. The study offers a way to separate those routes and then intervene on the one associated with abstract relations.
Each task presents a sample pair of objects and asks which candidate pair has the same relation. One candidate can preserve the relation while changing the objects; another can preserve surface appearance while changing the relation. The conflict exposes whether the model follows identity or structure.
Across GPT, Claude, Gemini, Qwen3.5, Gemma 4 and InternVL3 families, four changes shifted answers toward relations: a more capable model tier, a larger model, fewer objects in the scene and less noise attached to each object. The authors compare this progression with the “relational shift” observed in human development, but the resemblance is behavioral rather than proof of a shared cognitive mechanism.
The internal analysis found the two signals at different depths. Representations in earlier layers grouped examples by object-level features, while later layers grouped them by the abstract relation. Causal mediation tests then identified attention heads whose activity contributed to relation-based answers.
An intervention connects the circuit to another task
The strongest result comes from disabling the heads associated with relational matching. The authors report that this damages performance on ARC-AGI-1 more than disabling randomly selected heads. That transfer matters because the heads were found on the psychological matching task, rather than chosen directly to explain ARC results.
The evidence still has a narrow boundary. A set of heads can contribute to two tasks without constituting a general-purpose reasoning module. The stimuli are deliberately simple, and the paper does not establish that the same competition controls recognition in photographs, charts or long multimodal conversations.
The study also relies on two different levels of access. Behavioral results can include proprietary API models, but layer-by-layer representation and ablation require open weights. The mechanism claim therefore rests on the open models examined internally, while the broader family comparison is behavioral.
The parametric design is an advantage for reproduction. Object count and visual noise can be changed separately, letting another team test whether the layer split survives new shapes, relations and model families. The abstract page does not list a public code repository, so reproducing the full intervention will require more than downloading a benchmark.
The paper was submitted to arXiv on 6 October 2026 and has not been peer reviewed. Its useful contribution is a falsifiable bridge between a classic cognitive task and internal model circuitry: relation-following behavior should weaken when the identified late pathway is disrupted, while object matching should remain comparatively intact.
Verification
| Claim | Label | Primary source | Independent check |
|---|---|---|---|
| The study tests relational matching across frontier API and open vision-language models | VENDOR-REPORTED | preprint | none; author evaluation |
| Scale, capability, fewer objects and less noise shift models toward relation matching | VENDOR-REPORTED | preprint | none |
| Early layers encode object similarity while later layers encode abstract relations | VENDOR-REPORTED | preprint | none |
| Ablating relation-associated heads harms ARC-AGI-1 more than random-head ablation | VENDOR-REPORTED | preprint | none |
| Behavioral resemblance does not prove a shared human mechanism | ANALYSIS | preprint | inference from the study design |
| The preprint was submitted on 6 October 2026 | VERIFIED | arXiv record | arXiv metadata |