Determine whether VLM training favors common visual configurations

Determine whether vision-language-model training distributions systematically favor common object poses and spatial layouts, thereby creating configuration priors that bias object-direction and spatial-relation predictions in the absence of direct visual evidence.

Background

The paper evaluates object-direction and spatial-relation tasks in which the visually depicted orientation or arrangement can conflict with common poses and layouts. For these tasks, the authors use object identity as a proxy cue because no single prior answer is defined, and they observe behavior consistent with preferences for common configurations.

However, the source datasets do not provide model-specific training-frequency information, and the observed source-label distributions cannot establish what configurations the evaluated vision-LLMs encountered during training. The authors therefore formulate the existence of analogous training-induced configuration priors as a hypothesis rather than a demonstrated fact. Resolving this problem would require determining whether the training distributions of vision-LLMs systematically encode such preferences and quantifying their effect on visual predictions.

References

We hypothesize that VLM training similarly favors common configurations, but model-specific training frequencies are unavailable and cannot be inferred from these source counts.

— In-Context Learning Can Help Vision-Language Models Overcome Training Priors  (2610.12942 - Wang et al., 9 Oct 2026) in Appendix, Section “Training Priors for Object Direction and Spatial Relation” (Appendix configuration-priors diagnostics)