Study separates where cross-modal disagreement lives in omni models
An arXiv paper reports that the representation predicting when an omni model gives different answers across modalities is not the one whose perturbation changes those answers.
In plain words
Omni models handle audio, images, and text in one system, but sometimes answer the same question differently depending on the input type. The paper's authors found a small, reproducible pattern in the model's internal signals that predicts when such disagreement will happen. However, changing that pattern does not change the model's answers, while changing a different part does.
What was announced
The authors estimate a low-dimensional Answer Subspace from answer-position representations and use it to separate answer-related components from the rest of the representation. Across three omni models and two multimodal reasoning tasks, they report that information predictive of cross-modal inconsistency sits mainly outside this subspace, concentrated in a reproducible two-dimensional representation not explained by a fixed modality offset.
Matched interventions showed distinct roles: perturbing the Answer Subspace substantially changed predictions, while perturbing the disagreement representation left them almost unchanged, similar to a random subspace of the same size. The authors also report that disagreement can be induced by misaligning modality-specific answer representations, since perturbing the Answer Subspace for only one modality induced disagreement while applying the same perturbation to both largely preserved agreement. They conclude that linear decodability does not imply behavioral efficacy.
What it does not show
- The evidence is an abstract only; full experimental details, model names, and task definitions are not available in the supplied passages.
- The findings cover three omni models and two multimodal reasoning tasks as reported by the authors, with no independent confirmation.
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.