# Mosaic learns sparse image-text latents for VLMs

Digest AI · Research · published 2026-10-14T04:00:00Z

Canonical: https://digestai.news/story/mosaic-learns-sparse-image-text-latents-for-vlms

## Summary

Researchers introduced MOSAIC, a new method to understand vision-language models (VLMs). It creates a shared sparse latent space between images and text concepts. This enables localized influence attribution across the model's training data without degrading performance. MOSAIC was tested on Qwen and Gemma for tasks like captioning, X-ray understanding, and visual QA. The method shows how VLM predictions are influenced by specific image regions and text phrases, improving localization accuracy compared to existing methods.

## Key points

- MOSAIC learns sparse shared latent space between images and text
- Localized influence attribution without degrading model performance
- Improves localization accuracy over existing methods

## Why it matters

This method can help explain VLM predictions by showing where specific image regions and text phrases contribute to the model's decisions, which is crucial for understanding AI systems.

## Sources

1. [Cross-Modal Latent Influence: Patch-Level Training Sample Attribution for Vision-Language Models](https://arxiv.org/abs/2610.14190) (arXiv cs.AI, 2026-10-14, primary source)

## Cite

Digest AI, "Mosaic learns sparse image-text latents for VLMs", 14 October 2026, https://digestai.news/story/mosaic-learns-sparse-image-text-latents-for-vlms

---

Written by Digest AI's editorial model from the linked sources; the sources are the record. Headlines, digests and key points are written by Digest AI and may be quoted with a link to the story page. Linked articles belong to their publishers. Terms: https://digestai.news/terms#reuse
JSON: https://digestai.news/story/mosaic-learns-sparse-image-text-latents-for-vlms.json