Vision-language models are usually built by taking a language model and fine-tuning it to attend to visual features while optimizing a visual question-answering objective. A new arXiv preprint asks a more basic question: does this alignment actually change how the language model represents text? The answer implied by the paper's title is that it often does not — the model can fail to use its visual representations when working with text alone.
The authors describe this as a "failure of imagination," and they report that self-distillation can help. Rather than relying solely on the standard fine-tuning signal, their approach encourages the model to make better use of the visual knowledge it already has. The abstract stops short of giving full method details, but the central claim is clear: current alignment techniques are not enough on their own.
Because only one source is available here, there are no independent results to compare or contrast with these findings. The paper is a preprint, so the conclusions should be treated as provisional until they undergo peer review and replication. Still, the work points to a concrete limitation in how vision-language models are trained today. Source: arXiv:2610.25470.