Researchers have developed a method designed to show not only which training examples influenced a vision-language model’s prediction, but also which concepts, words and image regions carried that influence. The framework, called MOSAIC, is described in an arXiv preprint submitted October 9 by Laura Yao, Shixing Yu, Jie Peng, Promit Ghosal, Kyra Gan and Tianlong Chen. The paper is under review, so its findings should be treated as research results rather than established practice.

Most existing explanation tools answer only part of the question. Visual-attribution techniques can identify pixels or patches associated with a prediction, while training-data attribution can rank examples that appear to have shaped it. According to the paper, those approaches generally do not connect a particular training example to the exact multimodal evidence through which it matters. MOSAIC is intended to build that chain: prediction, training example, internal concept, and finally the relevant phrase or image region.

The method inserts a sparse autoencoder into an intermediate layer of a vision-language model. That component turns dense internal activity into a smaller set of active features, or latents. MOSAIC uses one shared dictionary for image and text representations, adds constraints that encourage visual features to occupy coherent regions instead of scattered patches, and aligns phrases with corresponding regions. An approximate influence calculation then estimates how individual training examples affect a target prediction and decomposes that effect across the learned features.

Dense image and language signals become a sparse set of aligned visual concepts.
The method learns shared image-text features while encouraging visual activations to stay concentrated in coherent regions.

The researchers tested the framework with two open-weight model families, Qwen and Gemma, across three settings. VizWiz-Captions evaluates descriptions of challenging real-world images, CheXpert covers classification of chest X-ray findings, and GQA tests visual question answering about objects, attributes and relationships. For each task, the team adapted the underlying model, froze it, and trained the sparse autoencoder on its intermediate representations.

A central requirement was preserving the behavior being explained. The paper reports only marginal performance changes after the sparse autoencoder was inserted across the tested tasks and models. At the same time, the learned features became more spatially concentrated. On GQA with Qwen, for example, the median share of image patches occupied by a feature fell from 0.688 with the baseline sparse autoencoder to 0.120 with the structured method.

The authors also tested whether the features ranked as influential actually mattered to the output. On GQA with Gemma, removing the top 256 influence-ranked features changed the model’s prediction in 38.3 percent of examples, compared with 8.3 percent when features were ranked by activation strength. In a complementary test, keeping only 64 influence-ranked features recovered the original answer for 75 percent of examples, versus 48.3 percent for activation-ranked features. The paper reports similar overall patterns on the captioning and chest X-ray tasks.

Removing visual evidence changes the training examples supporting an AI answer.
The experiments found that visual input sharply reordered influential training examples while much of the internal concept structure stayed more stable.

Another experiment examined what happens when the image is removed. Across six model-and-dataset combinations, rankings of influential internal features remained moderately to strongly related between the multimodal and text-only conditions, with reported Spearman correlations from 0.543 to 0.795. Rankings of influential training examples changed far more sharply, producing correlations from negative 0.797 to negative 0.564. The authors interpret that contrast as evidence that visual input can reorganize which training examples support an answer even when much of the conceptual structure remains comparatively stable.

The resulting hierarchy can also expose a grounded path to a mistake. In examples discussed by the researchers, correct answers were supported by influential examples and concepts aligned with the relevant visual evidence. Incorrect answers could instead be traced to a visually salient but mistaken concept. That distinction could make the approach useful for auditing failures, although the study does not demonstrate a deployed diagnostic system.

The paper identifies several limits. The explanations depend on the quality of the sparse representation, whose features may remain partly entangled and may change with the selected model layer or dictionary size. Phrase-to-region alignment uses the model’s existing attention as a correspondence signal, which can be incomplete or noisy. The influence calculation is also an approximation of what would happen under full retraining, and the experiments cover two relatively small model families rather than the largest vision-language systems.

MOSAIC therefore does not make a multimodal model fully transparent. Its contribution is narrower and potentially practical: it offers a way to inspect where a prediction’s training influence came from and what visible or linguistic evidence transmitted it. If the approach holds up at larger scales and under further validation, it could help researchers investigate data problems, hallucinations and model edits with more precision than either image heat maps or example rankings provide alone.