Lost in Embeddings: Information Loss in Vision-Language Models

TL;DR

This work addresses information loss in vision-language model connectors by introducing two complementary metrics: the $k$-NN overlap ratio (KNOR) to quantify geometric preservation and a patch-level embedding reconstruction to localize information loss. The authors evaluate three open-weight connector-based VLMs across six diverse datasets, revealing substantial distortion of local geometry (40–60% NN divergence) and meaningful patch-wise losses that correlate with captioning and visually grounded QA performance. They also show that linear alignment via Procrustes fails to reconcile pre- and post-projection spaces, underscoring the need for non-linear techniques or reconstruction-based regularization. The findings provide actionable insights for designing more faithful connectors and guiding future research on robust multimodal fusion.

Abstract

Vision--language models (VLMs) often process visual inputs through a pretrained vision encoder, followed by a projection into the language model's embedding space via a connector component. While crucial for modality fusion, the potential information loss induced by this projection step and its direct impact on model capabilities remain understudied. We introduce two complementary approaches to examine and quantify this loss by analyzing the latent representation space. First, we evaluate semantic information preservation by analyzing changes in k-nearest neighbor relationships between image representations, before and after projection. Second, we directly measure information loss by reconstructing visual embeddings from the projected representation, localizing loss at an image patch level. Experiments reveal that connectors substantially distort the local geometry of visual representations, with k-nearest neighbors diverging by 40--60\% post-projection, correlating with degradation in retrieval performance. The patch-level embedding reconstruction provides interpretable insights for model behavior on visually grounded question-answering tasks, finding that areas of high information loss reliably predict instances where models struggle.

Figures (11)

• Figure 1: Visualization of patch-wise information loss in the embeddings explains the incorrect predicted answer in VizWiz Grounding VQA. For the question "What is the fifth number?", LLaVA incorrectly predicted "18". Figure \ref{['fig:heatmap']} display the difference between the $L^2$ norm of the original and the reconstructed patch embeddings. Blue regions indicate where original embeddings have larger norms than predicted embeddings, while red regions show where predicted embeddings have larger norms. The top 10 high-loss patches are marked by yellow squares. Figure \ref{['fig:overlay']} shows high loss occurring in several answer-relevant patches contribute to the incorrect prediction.
• Figure 2: The $k$-nearest neighbors overlap ratio measures the overlap of an image's neighbors before and after projection. In this example, with $k=3$, the overlap ratio is 0.67 because two out of the three nearest neighbors are identical in both representation spaces.
• Figure 3: Neighborhood overlap ratios across three datasets: SeedBench validation, a 10,000-sample subset of VQAv2 validation, and Vizwiz grounding VQA validation. Analysis using 10, 50, and 100 nearest neighbors shows overlap ratios below 0.62 for all models, suggesting connectors poorly preserve geometric relationships and neighbor rankings for the visual representations.
• Figure 4: Comparison of five nearest neighbors searched with pre-projection (top) and post-projection (bottom) embeddings using different models. The first image in each row is the query image, followed by its nearest neighbors. For Qwen2.5-VL, despite a low neighborhood overlap ratio, post-projection embeddings retrieve more semantically similar images.
• Figure 5: Correlation between reconstruction loss and question-answering accuracy on the VizWiz grounding VQA task. For LLaVA and Idefics2, all correlations have a $p$-value < 5.0e-5, indicating statistically significant relationships, whereas no clear correlation is observed for Qwen2.5-VL. The reconstruction loss occurs in both answer-relevant and irrelevant patches. Loss in relevant patches negatively affects performance of LLaVA and Idefics2. "Norm" represents differences between the $L^2$ norm of the embeddings.