arXiv AI

CoPre-MoE: Consolidating and Preserving MoE for Multimodal Continual Instruction Tuning

arXiv Computation and Language
Aug 28

CRAM: Centroid-Routing and Adaptive MoE for Multimodal Continual Instruction Tuning

CRAM (Centroid‑Routing and Adaptive MoE) is a method for Multimodal Continual Instruction Tuning that isolates task‑specific patterns into independent modules to reduce catastrophic forgetting. It uses adaptive‑rank instantiation to allocate only the necessary parameters for new tasks, and centroid‑guided routing with an orthogonality penalty to reuse existing experts while preventing interference. Experiments on diverse benchmarks show CRAM outperforms existing approaches.

By Jun-Tao Tang, Zhen-Hao Xie, Yu-Cheng Shi, Da-Wei Zhou
arXiv Machine Learning
Jun 2

ProtoAda: Prototype-Guided Adaptive Adapter Expansion and Geometric Consolidation for Multimodal Continual Instruction Tuning

arXiv:2606. 02576v1 Announce Type: cross Abstract: Multimodal Large Language Models (MLLMs) achieve strong performance through instruction tuning, but real-world deployment requires them to continually acquire new vision-language capabilities, making Multimodal Continual Instruction Tuning (MCIT) essential.

By Yu-Cheng Shi, Zhen-Hao Xie, Jun-Tao Tang, Da-Wei Zhou
arXiv AI
Oct 21

Dynamic Cross-Modal Prompt Generation for Multimodal Continual Instruction Tuning

Dynamic Cross-Modal Prompt Generation (DRAPE) is a prompt-learning framework designed for Multimodal Continual Instruction Tuning (MCIT). It generates continuous, instance-specific soft prompts by cross-attending textual instructions with visual patch features, rather than selecting from a fixed prompt pool. DRAPE also incorporates null‑space gradient projection to reduce forgetting and uses CLIP‑based prototype routing for task‑label‑free inference, achieving state‑of‑the‑art results on MCIT benchmarks.

By Tao Hu, Da-Wei Zhou
arXiv AI
Oct 16

CART: Compatible Routing for Multimodal Continual Instruction Tuning

The paper introduces CART, a method for Multimodal Continual Instruction Tuning that coordinates adaptation between a projector and a language model by using task‑specific experts and weighted mixtures. CART jointly infers expert weights based on relevance and cross‑module compatibility, and it controls model growth through relevance‑guided expert consolidation. Experiments show that CART achieves state‑of‑the‑art performance on multimodal continual learning tasks.

By Ling-Xiang Zhang, Tao Hu, Da-Wei Zhou
arXiv Machine Learning
Aug 27

SAME: Stabilized Mixture-of-Experts for Multimodal Continual Instruction Tuning

The paper introduces SAME (Stabilized Mixture-of-Experts) to address challenges in Multimodal Continual Instruction Tuning (MCIT) for large language models. SAME mitigates router drift by decomposing routing dynamics into orthogonal subspaces and updating only task-relevant directions, while preventing expert drift through curvature‑aware scaling that uses historical input covariance without rehearsal. The method also employs adaptive expert activation to freeze selected experts during training, reducing redundant computation and cross‑task interference, and demonstrates state‑of‑the‑art performance on a new long‑task‑sequence benchmark.

By Zhen-Hao Xie, Jun-Tao Tang, Yu-Cheng Shi, Han-Jia Ye, De-Chuan Zhan, Da-Wei Zhou
arXiv Machine Learning
Oct 12

Monarch-MoE: Structured Experts for Multimodal Continual Instruction Tuning

Monarch-MoE introduces structured experts for multimodal continual instruction tuning by replacing dense LoRA factors with block‑diagonal ones linked by a fixed permutation, enabling blockwise gradient isolation. This design reduces adapter parameters by up to 75% and 50% on two benchmarks while keeping first‑task forgetting below 0.52 and 0.10 percentage points, respectively. The method is optimized for memory‑limited hardware, allowing fine‑tuning and inference on a single 24 GB RTX 4090 with lower memory usage and competitive runtime.

By Xi Zhang, Boyu Diao, Zesen Chen, Ruiqi Liu, Hangda Liu, Yongjun Xu
arXiv AI
Oct 19

COMET: Decoupled Distillation, Routing, and Capacity Control for Task-Agnostic Continual Vision--Language Learning

COMET is a task‑agnostic continual learning framework for vision‑language models that addresses modal‑consistency drift, representation‑space interference, and per‑task capacity shortfall. It uses a frozen teacher for feature reference, Tri‑Affinity Distillation to preserve image‑image, text‑text, and image‑text geometry, and Student‑Aware Cross‑Expert Distillation to align new experts with prior ones, while a contextual bandit and reconstruction router manage capacity and inference. Experiments on the X‑TAIL benchmark show that each component improves continual learning performance, demonstrating that separating feature reference, expert shaping, routing, and capacity control enables effective multimodal adaptation.

By Seungyoon Woo, Keonhee Park, Gunhee Kim
arXiv Computer Vision
Oct 19

Mitigating Task Objective Conflicts for Continual Learning in Unified Multimodal Models

The paper introduces UniDecouple, a selective decoupling framework for continual learning in unified multimodal models that addresses Task Objective Conflict (TOC). By employing a Task-Aware Mixture-of-Experts (TA‑MoE) architecture, it retains shared attention layers while adding task‑specific feed‑forward network experts to reduce optimization conflicts. Hierarchical expert routing and hybrid collaboration further balance task specialization with cross‑task knowledge sharing, and a two‑stage training strategy enhances both multimodal understanding and image generation performance.

By Jiaxing Zhang, Hao Tang