Information: Join today and make your research activities more affordable! Technical workshop participation fees and annual registration fees are available at member rates.
Notice: [Important] Announcement of Changes to Registration Fee Payment and Manuscript Upload Procedures for IEICE Technical Meetings
IEICE Technical Committee Submission System
Conference Paper's Information		 Online Proceedings
[Sign in]
Tech. Rep. Archives
 Go Top Page Go Previous   [Japanese] / [English] 

Paper Abstract and Keywords
Presentation 2026-03-01 14:20
Do Vision-Language Models Actually Use Images? -- An Analysis Based on Output Text and Internal Confidence Dynamics --
Yuka Miyata, Chie Mizunuma (JWU), Noriyuki Okumura (MWU), Kimio Kuramitsu (JWU) NLC2025-19
Abstract (in Japanese) (See Japanese page) 
(in English) Vision-Language Models (VLMs) process information by integrating images and text; however, it remains difficult to quantitatively evaluate the extent to which their outputs are actually grounded in visual information.
In this study, we address the question of whether VLMs truly utilize images by focusing on changes in internal confidence caused by the presence or absence of visual input. We conduct a layer-wise analysis using a layer-specific internal confidence metric, $Delta C^{(l)}$. Two VLMs with projection-based architectures, sarashina2-vision-8B and Qwen3-VL-8B, are examined by measuring changes in output probability distributions across Transformer layers to investigate where visual information exerts the strongest influence.
Experimental results reveal differences between the models in terms of Peak Layer distributions and the magnitude of $Delta C$, suggesting that the location and manner of visual information integration vary across architectures. Furthermore, analysis of output diversity and length indicates that $Delta C$ captures changes in generation behavior associated with visual input. This study provides one perspective for analyzing visual information utilization in VLMs based on internal state dynamics.
Keyword (in Japanese) (See Japanese page) 
(in English) Vision-Language Model / Internal Confidence / Layer-wise Analysis / Vision-Language Fusion / / / /  
Reference Info. IEICE Tech. Rep., vol. 125, no. 366, NLC2025-19, pp. 40-45, Feb. 2026.
Paper # NLC2025-19 
Date of Issue 2026-02-21 (NLC) 
ISSN Online edition: ISSN 2432-6380
Copyright
and
reproduction		 All rights are reserved and no part of this publication may be reproduced or transmitted in any form or by any means, electronic or mechanical, including photocopy, recording, or any information storage and retrieval system, without permission in writing from the publisher. Notwithstanding, instructors are permitted to photocopy isolated articles for noncommercial classroom use without fee. (License No.: 10GA0019/12GB0052/13GB0056/17GB0034/18GB0034)
Download PDF NLC2025-19

Conference Information
Committee NLC  
Conference Date 2026-02-28 - 2026-03-01 
Place (in Japanese) (See Japanese page) 
Place (in English)  
Topics (in Japanese) (See Japanese page) 
Topics (in English)  
Paper Information
Registration To NLC 
Conference Code 2026-03-NLC 
Language Japanese 
Title (in Japanese) (See Japanese page) 
Sub Title (in Japanese) (See Japanese page) 
Title (in English) Do Vision-Language Models Actually Use Images? 
Sub Title (in English) An Analysis Based on Output Text and Internal Confidence Dynamics 
Keyword(1) Vision-Language Model  
Keyword(2) Internal Confidence  
Keyword(3) Layer-wise Analysis  
Keyword(4) Vision-Language Fusion  
Keyword(5)  
Keyword(6)  
Keyword(7)  
Keyword(8)  
1st Author's Name Yuka Miyata  
1st Author's Affiliation Japan Women's University (JWU)
2nd Author's Name Chie Mizunuma  
2nd Author's Affiliation Japan Women's University (JWU)
3rd Author's Name Noriyuki Okumura  
3rd Author's Affiliation Mukogawa Women's University (MWU)
4th Author's Name Kimio Kuramitsu  
4th Author's Affiliation Japan Women's University (JWU)
5th Author's Name  
5th Author's Affiliation ()
6th Author's Name  
6th Author's Affiliation ()
7th Author's Name  
7th Author's Affiliation ()
8th Author's Name  
8th Author's Affiliation ()
9th Author's Name  
9th Author's Affiliation ()
10th Author's Name  
10th Author's Affiliation ()
11th Author's Name  
11th Author's Affiliation ()
12th Author's Name  
12th Author's Affiliation ()
13th Author's Name  
13th Author's Affiliation ()
14th Author's Name  
14th Author's Affiliation ()
15th Author's Name  
15th Author's Affiliation ()
16th Author's Name  
16th Author's Affiliation ()
17th Author's Name  
17th Author's Affiliation ()
18th Author's Name  
18th Author's Affiliation ()
19th Author's Name  
19th Author's Affiliation ()
20th Author's Name  
20th Author's Affiliation ()
21st Author's Name  
21st Author's Affiliation ()
22nd Author's Name  
22nd Author's Affiliation ()
23rd Author's Name  
23rd Author's Affiliation ()
24th Author's Name  
24th Author's Affiliation ()
25th Author's Name  
25th Author's Affiliation ()
26th Author's Name / /
26th Author's Affiliation ()
()
27th Author's Name / /
27th Author's Affiliation ()
()
28th Author's Name / /
28th Author's Affiliation ()
()
29th Author's Name / /
29th Author's Affiliation ()
()
30th Author's Name / /
30th Author's Affiliation ()
()
31st Author's Name / /
31st Author's Affiliation ()
()
32nd Author's Name / /
32nd Author's Affiliation ()
()
33rd Author's Name / /
33rd Author's Affiliation ()
()
34th Author's Name / /
34th Author's Affiliation ()
()
35th Author's Name / /
35th Author's Affiliation ()
()
36th Author's Name / /
36th Author's Affiliation ()
()
Speaker Author-1 
Date Time 2026-03-01 14:20:00 
Presentation Time 25 minutes 
Registration for NLC 
Paper # NLC2025-19 
Volume (vol) vol.125 
Number (no) no.366 
Page pp.40-45 
#Pages 6 
Date of Issue 2026-02-21 (NLC) 

[Return to Top Page]

[Return to IEICE Web Page]
The Institute of Electronics, Information and Communication Engineers (IEICE), Japan