Where MLLMs Attend and What They Rely On: Explaining Autoregressive Token Generation

Fuente: arXiv
Saved in:
Bibliographic Details
Main Authors: Chen, Ruoyu, Guo, Xiaoqing, Liu, Kangwei, Liang, Siyuan, Liu, Shiming, Zhang, Qunli, Wang, Laiyuan, Zhang, Hua, Cao, Xiaochun
Format: Preprint
Published: 2025
Subjects:
Online Access:
Tags: Add Tag
No Tags, Be the first to tag this record!
_version_ 1866908895171051520
author Chen, Ruoyu
Guo, Xiaoqing
Liu, Kangwei
Liang, Siyuan
Liu, Shiming
Zhang, Qunli
Wang, Laiyuan
Zhang, Hua
Cao, Xiaochun
author_facet Chen, Ruoyu
Guo, Xiaoqing
Liu, Kangwei
Liang, Siyuan
Liu, Shiming
Zhang, Qunli
Wang, Laiyuan
Zhang, Hua
Cao, Xiaochun
contents Multimodal large language models (MLLMs) have demonstrated remarkable capabilities in aligning visual inputs with natural language outputs. Yet, the extent to which generated tokens depend on visual modalities remains poorly understood, limiting interpretability and reliability. In this work, we present EAGLE, a lightweight black-box framework for explaining autoregressive token generation in MLLMs. EAGLE attributes any selected tokens to compact perceptual regions while quantifying the relative influence of language priors and perceptual evidence. The framework introduces an objective function that unifies sufficiency (insight score) and indispensability (necessity score), optimized via greedy search over sparsified image regions for faithful and efficient attribution. Beyond spatial attribution, EAGLE performs modality-aware analysis that disentangles what tokens rely on, providing fine-grained interpretability of model decisions. Extensive experiments across open-source MLLMs show that EAGLE consistently outperforms existing methods in faithfulness, localization, and hallucination diagnosis, while requiring substantially less GPU memory. These results highlight its effectiveness and practicality for advancing the interpretability of MLLMs.
format Preprint
id arxiv_https___arxiv_org_abs_2509_22496
institution arXiv
publishDate 2025
record_format arxiv
spellingShingle Where MLLMs Attend and What They Rely On: Explaining Autoregressive Token Generation
Chen, Ruoyu
Guo, Xiaoqing
Liu, Kangwei
Liang, Siyuan
Liu, Shiming
Zhang, Qunli
Wang, Laiyuan
Zhang, Hua
Cao, Xiaochun
Computer Vision and Pattern Recognition
Multimodal large language models (MLLMs) have demonstrated remarkable capabilities in aligning visual inputs with natural language outputs. Yet, the extent to which generated tokens depend on visual modalities remains poorly understood, limiting interpretability and reliability. In this work, we present EAGLE, a lightweight black-box framework for explaining autoregressive token generation in MLLMs. EAGLE attributes any selected tokens to compact perceptual regions while quantifying the relative influence of language priors and perceptual evidence. The framework introduces an objective function that unifies sufficiency (insight score) and indispensability (necessity score), optimized via greedy search over sparsified image regions for faithful and efficient attribution. Beyond spatial attribution, EAGLE performs modality-aware analysis that disentangles what tokens rely on, providing fine-grained interpretability of model decisions. Extensive experiments across open-source MLLMs show that EAGLE consistently outperforms existing methods in faithfulness, localization, and hallucination diagnosis, while requiring substantially less GPU memory. These results highlight its effectiveness and practicality for advancing the interpretability of MLLMs.
title Where MLLMs Attend and What They Rely On: Explaining Autoregressive Token Generation
topic Computer Vision and Pattern Recognition
url https://arxiv.org/abs/2509.22496