Modality-Decoupled Online Recursive Editing
The paper introduces M-ORE, a modality-decoupled online recursive editor designed for multimodal large language models (MLLMs). This method addresses challenges faced by traditional text-only editors, such as cross-modal conflicts and long-horizon interference. Experimental results demonstrate that M-ORE enhances reliability, generality, and locality while maintaining efficiency in updates.
- ▪M-ORE is developed to improve online model editing for multimodal large language models.
- ▪The method uses a unified proximal-projection formulation for constant per-edit overhead.
- ▪Experiments show that M-ORE consistently outperforms strong baselines in various benchmarks.
Opening excerpt (first ~120 words) tap to expand
Computer Science > Machine Learning arXiv:2605.20273 (cs) [Submitted on 19 May 2026] Title:Modality-Decoupled Online Recursive Editing Authors:Siyuan Li, Youyuan Zhang, Fangming Liu, Jing Li View a PDF of the paper titled Modality-Decoupled Online Recursive Editing, by Siyuan Li and 3 other authors View PDF HTML (experimental) Abstract:Online model editing for multimodal large language models (MLLMs) requires assimilating a stream of corrections under tight compute and memory budgets. Yet editors developed for text-only LLMs often degrade on MLLMs: visually dominant activations skew the statistics that shape updates, causing cross-modal conflict, while sequential writes become entangled in a shared edit space and amplify long-horizon interference, causing inter-edit interference.
…
Excerpt limited to ~120 words for fair-use compliance. The full article is at arXiv cs.AI.
Discussion
0 commentsMore from arXiv cs.AI
