MARS: Harmonizing Multimodal Convergence via Adaptive Rank Search
This addresses the challenge of efficient and automated fine-tuning for multimodal AI systems, though it is incremental as it builds on existing LoRA methods.
The paper tackles the problem of imbalanced training dynamics in fine-tuning Multimodal Large Language Models with parameter-efficient methods like LoRA, which causes suboptimal accuracy due to negative interference, and introduces MARS to discover optimal rank pairs that balance dynamics and maximize performance, outperforming baseline methods.
Fine-tuning Multimodal Large Language Models (MLLMs) with parameter-efficient methods like Low-Rank Adaptation (LoRA) is crucial for task adaptation. However, imbalanced training dynamics across modalities often lead to suboptimal accuracy due to negative interference, a challenge typically addressed with inefficient heuristic methods such as manually tuning separate learning rates. To overcome this, we introduce MARS (Multimodal Adaptive Rank Search), an approach to discover optimal rank pairs that balance training dynamics while maximizing performance. Our key innovation, a proposed framework of dual scaling laws, enables this search: one law models module-specific convergence time to prune the search space to candidates with aligned dynamics, while the other predicts final task performance to select the optimal pair from the pruned set. By re-purposing the LoRA rank as a controller for modality-specific convergence speed, MARS outperforms baseline methods and provides a robust, automated strategy for optimizing MLLM fine-tuning.