Research map / Vision-language and multimodal / Multimodal LLMs

Modality-Decoupled Online Recursive Editing

ICML 2026 · Vision-language and multimodal › Multimodal LLMs · approach: Unified any-modality language models · 0 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Proposes M-ORE, a modality-decoupled online recursive editor for MLLMs that keeps separate locality statistics for the text stack and visual projector and performs Sherman-Morrison updates in a frozen orthogonal low-rank write space, achieving constant per-edit overhead.

Problem

Online model editing of multimodal LLMs under a stream of corrections suffers from cross-modal conflict (visual activations dominate update statistics) and inter-edit interference (sequential edits entangle in a shared subspace), while existing editors either drift or incur edit-growing cost.

Approach

Derives a proximal-projection update with a locality quadratic form, restricts writes to a frozen orthonormal low-rank basis, keeps separate per-module (text vs. visual projector) steady-space preconditioners updated by Sherman-Morrison recursion, giving O(1) per-edit cost.

Key results

• On E-IC with LLaVA-v1.5 at 100 edits, M-ORE reaches 90.73 average vs 83.66 for LiveEdit.
• On BLIP2-OPT E-IC at 100 edits, reliability improves from 74.63 (LiveEdit) to 83.14 for M-ORE.
• Removing the frozen orthogonal basis drops LLaVA E-IC average by 18.10 points at 100 edits.

Surprising findings

• Removing masked pooling slightly improves the average score (+1.70 at 100 edits on E-IC), suggesting the rank-one sketch is not harmful.
• Editing the vision encoder and projector alone preserves text locality but lowers reliability and generality.

Limitations

• Cross-modal covariance is approximated with diagonal variance and layer-wise energy proxies.
• Some locality flips occur near fragile decoding boundaries and with visually similar multimodal inputs.

Experimental setup

Datasets: E-VQA E-IC NQ OK-VQA MME

Benchmarks: E-VQA E-IC MME

Metrics: reliability generality locality average score edit latency peak memory

Baselines: FT-L FT-M MEND AlphaEdit SERAC IKE LiveEdit

Builds on: ROME MEMIT AlphaEdit LiveEdit MEND

Closest papers by meaning

• Rethinking Residual Distribution in Locate-then-Edit Model Editing (NeurIPS 2025 · similarity 0.88)
• Towards Lifelong Model Editing via Simulating Ideal Editor (ICML 2025 · similarity 0.88)
• MEMOIR: Lifelong Model Editing with Minimal Overwrite and Informed Retention for LLMs (NeurIPS 2025 · similarity 0.88)
• Reinforced Lifelong Editing for Language Models (ICML 2025 · similarity 0.87)
• Lifelong Knowledge Editing for Vision Language Models with Low-Rank Mixture-of-Experts (CVPR 2025 · similarity 0.87)
• BalancEdit: Dynamically Balancing the Generality-Locality Trade-off in Multi-modal Model Editing (ICML 2025 · similarity 0.87)