AdaptVis: Spatial Understanding in Vision-Language Models Requires Adaptive Attention

Shiqi Chen, Tongyao Zhu, Ruochen Zhou, Siyang Gao, Juan Carlos Niebles, Mor Geva, Junxian He, Jiajun Wu, Manling Li

Keywords: Vision Language Models, Uncertainty, Mechanistic interpretability, Constrain Decoding, Spatial Understanding

TL;DR: We identify that the distribution of attention scores on images plays a crucial role in spatial reasoning tasks within VLLMs. To enhance performance, we propose an adaptive control method that dynamically adjusts these attention scores.

Abstract: Vision Large Language Models (VLLMs) often struggle with adequately attending to image information, leading to significant hallucinations across various domains, especially on spatial reasoning tasks. In this study, we analyze the attention behavior of VLLMs in spatial reasoning Question-Answering (QA) tasks from a mechanism interpretability view. By visualizing the crucial areas of an image that receive the highest attention scores in the intermediate layers, we identify an interesting pattern: failures often correspond to attention being misallocated to irrelevant objects within the image. Moreover, the attention patterns exhibit large differences between familiar and unfamiliar spatial relationships. Motivated by this observation, we further explore the feasibility of adaptively adjusting the attention scores during the inference process based on the confidence score. Our experiments on spatial reasoning benchmarks including WhatsUp and VSR demonstrate that our decoding methods yield promising results, e.g., achieving up to a 50-point improvement on the WhatsUp benchmark with negligible additional computation cost.

Primary Area: foundation or frontier models, including LLMs

Code Of Ethics: I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics.

Submission Guidelines: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2025/AuthorGuide.

Reciprocal Reviewing: I understand the reciprocal reviewing requirement as described on https://iclr.cc/Conferences/2025/CallForPapers. If none of the authors are registered as a reviewer, it may result in a desk rejection at the discretion of the program chairs. To request an exception, please complete this form at https://forms.gle/Huojr6VjkFxiQsUp6.

Anonymous Url: I certify that there is no URL (e.g., github page) that could be used to find authors’ identity.

No Acknowledgement Section: I certify that there is no acknowledgement section in this submission for double blind review.

Submission Number: 4162