Intuition vectors from frozen vision encoders steer diffusion reasoning on ARC tasks
Researchers report that simple vector differences between image embeddings can guide a diffusion transformer on abstract reasoning, outperforming demonstration conditioning on held-out ARC tasks while using far less test-time compute.
In plain words
The authors of an arXiv abstract say that subtracting one image embedding from another, which they call an intuition vector, captures the transformation shown in a task's examples. They then use that vector to steer a diffusion model, a generative image model, toward the right answer. On new ARC tasks, this method reportedly beat feeding the model full examples while using far less computing time.
What was announced
The study, described in the paper's abstract, examines whether abstract reasoning can rest on the relational geometry of pretrained visual representations alone, without language-based inference or task-specific representation learning. Working with frozen self-supervised vision encoders, the authors found that vector differences between image embeddings, which they name intuition vectors, capture transformations that generalize across instances of the same task. These vectors reportedly distinguish among 794 abstract reasoning tasks with up to 87.2 percent accuracy on Bongard problems and encode systematic relationships between input-output pairs in ARC tasks.
To test whether this relational structure can steer generative reasoning, the team conditioned a diffusion transformer operating in a pretrained visual latent space. On held-out ARC tasks, they report that intuition-vector conditioning outperformed conditioning on complete demonstrations while requiring nearly six times less test-time compute, and that it supports zero-shot inference. The authors present this as evidence that pretrained visual representations can serve as a transferable substrate for abstract reasoning, though the claims come from the abstract alone and await independent evaluation.
What it does not show
- The evidence is limited to the paper's abstract; full experimental details, methods and results are not available for verification.
- Reported accuracy and compute figures are the authors' own claims and have not been independently confirmed.
For security and IT teams
- Scope
- The reported results cover Bongard problems and ARC tasks using frozen self-supervised vision encoders; the abstract does not describe performance on other benchmarks or in deployed systems.
Sources
Source-support review completed 19 Nov 2026. This story was drafted with AI assistance and checked against retained source passages. Automated checking is not independent verification.
Coverage: arXiv: abstract.
Updates and corrections
No corrections so far.
First published 19 Nov 2026
New to the subject? AI from the Ground Up starts from the beginning.