Separate geometric information from visual-appearance transfer in reference conditioning

Establish how much of the accuracy gain from a drawn reference dial is attributable to its hand-angle geometry rather than to changes in the generated dials’ visual appearance.

Background

The drawn reference dial improved the proportion of images in which all three clocks showed the requested times, but some models also copied aspects of the reference’s appearance, producing flatter or more diagram-like dials. Consequently, the comparison does not isolate the causal effect of geometric information.

The authors identify concrete controls that could resolve this issue: a diagram without hands, a reference rendered in the photographic style of the target object, or a digit-based prompt accompanied by a reference. Determining the separate contributions of geometry and appearance would clarify the mechanism and improve the design of visual-conditioning interventions.

References

The object is therefore not perfectly fixed in D, and part of the gain may come from a simpler dial being easier to draw, or to read, correctly. D estimates the combined effect of reference conditioning; separating the geometric information from the appearance would need further controls, such as the same diagram without hands, a reference drawn in the photographic style of the target object, or the time in digits plus a reference. We leave them to a follow-up study.

— It's Always 10:10: Reference Images Break a Bias That Prompts Only Dent  (2610.11320 - Cazzaniga, 8 Oct 2026) in Discussion, paragraph “What the reference also transfers”