Vision language models fail to compare matching points across images

Seen but Unsaid: Locating and Repairing Cross-Image Comparison in Vision-Language Models

Computer Vision and Pattern RecognitionComputation and LanguageMachine Learning

Summary

Vision-language models sometimes struggle to recognize the same point or object in two different images, even though they have the information internally. The authors identified that specific parts of the models, called attention heads, carry out this comparison but the answer does not get properly read out. They found a way to fix this by adjusting only a tiny fraction of the model’s parameters, which improves accuracy on several tests. This fix helps the models better connect related objects across multiple images.

What this means in practice

  • For computer vision engineers: Improve model accuracy in matching points or objects between images for tasks like object tracking or image alignment by fine-tuning specific attention heads.
  • For multimodal ai developers: Enhance cross-image reasoning capabilities in vision-language applications such as image captioning or visual question answering by repairing attention pathways.

Authors

Zhaoning Zhang

Abstract

Vision-language models (VLMs) often fail to match a point or an object across two images, even when the answer can be decoded from their own hidden states. We trace this failure from the visual tokens to the answer letter, in Qwen3-VL (2B-32B), InternVL3.5-8B, and Gemma-3-12B, on synthetic shape matching, same-instance point correspondence (HPatches), and semantic correspondence (SPair-71k), and find two distinct failure points, a small set of attention heads that carries the comparison, and a repair that touches 0.02% of the parameters. Two failure points: exact copies are compared but not read out. An answer-position probe exceeds the calibrated output by 15 to 25 points in every family, for nameable and unnameable shapes alike, and steering along the probe's direction moves no more answers than random directions. Semantic correspondences are present in the visual tokens but never reach the answer. One pathway: redistributing the cross-image attention of 16-32 heads, with its mass preserved, removes 17-22 points of synthetic and 15-32 points of HPatches accuracy, against at most 7 for random or attention-matched heads. A lever that training finds on its own: training only these heads' query and key projections raises SPair-71k accuracy by 28.3 points (Qwen3-VL-8B) and 25.1 points (InternVL3.5-8B), beats both layer-matched and attention-matched heads, and transfers to HPatches and BLINK. An unconstrained all-layer adapter routes 21-51% of its gain through the attention of the same heads, against at most 7% for random heads, and does so by changing what they read (83-86% of the effect) rather than their weights. Every decision rule was frozen before its run, and every reported number is recomputed by independent verification scripts.