Han, Xiaoqi, Li, Ru, Yi, Ran, Tan, Hongye, Liang, Zhuomin, Gutierrez Basulto, Victor  ORCID: https://orcid.org/0000-0002-6117-5459 and Pan, Jeff Z.
2025.
Uncovering and mitigating transient blindness in multimodal model editing.
Presented at: 40th Annual AAAI Conference on Artificial Intelligence (AAAI'26),
Singapore,
20-27 January 2026.
|
Preview |
PDF
- Presentation
Download (1MB) | Preview |
Abstract
Multimodal Model Editing (MMED) aims to correct erroneous knowledge in multimodal models. Existing evaluation methods, adapted from textual model editing, overstate success by relying on low-similarity or random inputs, obscure overfitting. We propose a comprehensive locality evaluation framework, covering three key dimensions: random-image locality, no-image locality, and consistent-image locality, operationalized through seven distinct data types, enabling a detailed and structured analysis of multimodal edits. We introduce De-VQA, a dynamic evaluation for visual question answering, uncovering a phenomenon we term transient blindness, overfitting to edit similar text while ignoring visuals. Token analysis shows edits disproportionately affect textual tokens. We propose locality-aware adversarial losses to balance cross-modal representations. Empirical results demonstrate that our approach consistently outperforms existing baselines, reducing transient blindness and improving locality by 17% on average.
| Item Type: | Conference or Workshop Item (Paper) |
|---|---|
| Status: | Unpublished |
| Schools: | Schools > Computer Science & Informatics |
| Related URLs: | |
| Date of First Compliant Deposit: | 27 November 2025 |
| Date of Acceptance: | 7 November 2025 |
| Last Modified: | 28 Nov 2025 16:00 |
| URI: | https://orca.cardiff.ac.uk/id/eprint/182717 |
Actions (repository staff only)
 |
Edit Item |
Download Statistics
Download Statistics