StrLoRA: Towards Streaming Continual Visual Instruction Tuning for MLLMs

Unknown Date

AI Summary

This work addresses the limitations of existing continual visual instruction tuning methods, which are largely confined to task-incremental settings and struggle with dynamically interleaved, continuously arriving multimodal data streams in real-world scenarios. To bridge this gap, we introduce Streamed Continual Visual Instruction Tuning (StrCVIT)—a novel setting that better reflects realistic deployment conditions—and establish a corresponding benchmark. We propose StrLoRA, a framework featuring a two-stage task-aware expert routing mechanism that enables sparse activation and token-level cross-modal weighting. Additionally, we incorporate a routing stability regularizer based on exponential moving average to mitigate catastrophic forgetting. Experimental results demonstrate that StrLoRA significantly outperforms current approaches on the StrCVIT benchmark, effectively enabling multimodal large language models to continually learn and enhance their capabilities under dynamic data streams.

Abstract

Continual Visual Instruction Tuning (CVIT) enables Multimodal Large Language Models to incrementally acquire new abilities. However, existing CVIT methods operate under a restrictive task-incremental setting, where each training phase corresponds to a single, predefined task. This does not reflect real-world conditions, where data arrives as a continuous stream of interleaved and dynamically evolving tasks. To bridge this gap, we introduce Streaming CVIT (StrCVIT), a more general and realistic setting where models learn from a stream of data chunks containing a dynamic mixture of tasks. In StrCVIT, a model must simultaneously acquire new abilities, reinforce recurring abilities, and mitigate forgetting. Existing CVIT methods fail here as they cannot reliably distinguish or adapt to the heterogeneous task samples within each chunk. We therefore propose StrLoRA, a regularized two-stage expert routing framework. StrLoRA first performs task-aware expert selection using the textual instruction to activate a sparse subset of relevant experts, reducing cross-task interference. It then applies token-wise expert weighting within this subset, where contribution weights are computed via cross-modal attention between local visual tokens and the global instruction representation. To maintain stability across the non-stationary stream, a routing-stability regularization aligns current routing distributions with a historical exponential moving average reference. Extensive experiments on a newly developed StrCVIT benchmark show that StrLoRA substantially outperforms existing methods, effectively enhancing model's abilities from continuously evolving data streams. The code is available at https://github.com/chanceche/StrCVIT.

Problem

• Streaming Continual Learning
• Multimodal Large Language Models
• Visual Instruction Tuning
• Task Interference
• Catastrophic Forgetting

Innovation

• Streaming Continual Learning
• Multimodal Large Language Models
• LoRA
• Expert Routing
• Cross-modal Attention