A Rendered Passage Is Not a Picture: Four Things Called Modality Preference in Multimodal Knowledge Conflict, Why the Reported Reversal of Text Dominance Changes the Channel but Not the Content, and the Reliability-Matched Test That Would Settle It

@misc{mahendrakar2026rendered,
  author       = {Mahendrakar, Pranay},
  title        = {{A Rendered Passage Is Not a Picture: Four Things Called Modality Preference in Multimodal Knowledge Conflict, Why the Reported Reversal of Text Dominance Changes the Channel but Not the Content, and the Reliability-Matched Test That Would Settle It}},
  year         = {2026},
  month        = nov,
  date         = {2026-11-03},
  publisher    = {Life of Research},
  howpublished = {Preprint},
  doi          = {10.5281/zenodo.23919442},
  url          = {https://lifeofresearch.com/preprints/185},
  abstract     = {(c) 2026 Pranay Mahendrakar. Licensed under CC BY 4.0. Life of Research A recurring finding about multimodal language models is that they trust text over images when the two disagree. A 2026 study of thirteen models reports the reverse for one kind of evidence: a passage that contradicts the model's memory is adopted more often when it is shown as a rendered image than when it is typed, and when a rendered passage and a typed passage contradict each other the winning channel changes with order, model and dataset. In this analysis the author reviews and synthesises published results to show that the reversal and the earlier finding are not about the same quantity. "Modality preference" names at least four measurements: how attention is shared across token types, which source a model follows when two sources disagree, how far a named target modality is disturbed by the other one, and whether the same content gives the same answer in two channels. The text-dominance studies mostly vary content type, pitting a written proposition against a perceptual scene; the 2026 study varies the channel while holding the proposition fixed. Typographic attacks, in which words inside an image override what the image shows, indicate that what has been called text dominance travels with propositional content rather than with the token channel. Re-reading the 2026 study's own tables, the author shows that its order effect is structured rather than arbitrary (in 11 of 13 models on one dataset the image gains when it is placed second), that most single-evidence gaps are no larger in magnitude than the study's same-modality controls (though the controls lean negative, and on a signed comparison seven of thirteen NQ-Swap gaps exceed all of them), that its normalized gap hides that, in the cells where the rates can be recovered, two proprietary models are not counted as following either passage on most items, and that for the five models, from three families, it shares with an earlier benchmark the sign of the cross-modal preference agrees on both datasets. Human cue-combination research resolved a similar dispute by showing that dominance follows relative reliability. The paper states a reliability-matched protocol as an algorithm and specifies what a claim of modality bias should report.},
  keywords     = {multimodal large language models, knowledge conflict, modality preference, text dominance, text-as-image, cue combination, position bias, evaluation methodology},
  language     = {en},
  copyright    = {Creative Commons Attribution 4.0}
}