Cross-Modal Latent Influence: Patch-Level Training Sample Attribution for Vision-Language Models
Explaining a vision--language model (VLM) prediction requires tracing training-example influence to the multimodal evidence that mediates it, yet existing attribution methods largely operate only at the sample level.
This article introduces MOSAIC, a method for improving the explainability of vision-language models by enabling patch-level training sample attribution, addressing limitations in existing attribution techniques that operate only at the sample level.
What happened
The article presents MOSAIC, a novel approach for attributing training examples to specific multimodal evidence in vision-language models, enhancing the explainability of model predictions by operating at the patch level rather than the sample level.
Why it matters
This development is significant for AI developers and researchers as it improves the transparency and interpretability of vision-language models, which are crucial for applications requiring explainable AI, such as medical imaging and complex decision-making systems.
- MOSAIC improves the explainability of vision-language models by enabling patch-level attribution.
- The method's sparse and decorrelated latent space allows for more precise localization of model influences.
- Evaluations show that MOSAIC maintains task performance while enhancing interpretability.
The bigger picture
MOSAIC enhances the ability to understand and trust vision-language models by providing more precise and localized attribution of training influences, which can lead to better model debugging, improved transparency, and more reliable applications in critical domains.
The article presents MOSAIC as a significant advancement in the explainability of vision-language models, offering a method to attribute training influences at the patch level, thereby enhancing transparency and reliability in AI systems.
Technical details
- MOSAIC (Multimodal Orthogonalized Sparse Autoencoders for Influence Characterization)
- Shared sparse image-text latent space with sparsity, contiguity, alignment, and decorrelation properties
- EK-FAC (Expected Kalman Filtering Approximation for Covariance) for decomposing training-example influence
- Evaluation on Qwen and Gemma models across captioning, chest X-ray understanding, and visual question answering tasks
FAQ
What is MOSAIC and what problem does it solve?
MOSAIC is a method for improving the explainability of vision-language models by enabling patch-level training sample attribution. It addresses the limitation of existing methods that operate only at the sample level, allowing for more precise localization of model influences.
How does MOSAIC work?
MOSAIC learns a shared sparse image-text latent space with specific properties and uses EK-FAC to decompose training-example influence through these concepts, localizing contributions to specific text phrases and image regions.
What are the key benefits of using MOSAIC?
MOSAIC provides more localized and coherent features without degrading task performance, enhances the transparency of vision-language models, and allows for better understanding of how training data influences model behavior.
This summary and analysis were generated with AI from the original article and may contain errors. Read the original for the full story.
Read the full story
Originally published by arXiv - cs.AI on Oct 14, 2026.
Comments
Sign in to join the discussion.