MLLM-CBench:A Comprehensive Benchmark for Continual Instruction Tuning of Multimodal LLMs with Chain-of-Thought Reasoning Analysis

Abstract: Multimodal LLMs (MLLMs) rely on continual instruction tuning to adapt to the evolving demands of real-world applications. However, progress in this area is hindered by the lack of rigorous and systematic benchmarks. To address this gap, we present MLLM-CTBench, a comprehensive evaluation benchmark with three key contributions: (1) Multidimensional Evaluation: We combine final answer accuracy with fine-grained CoT reasoning quality assessment, enabled by a specially trained CoT evaluator; (2) Comprehensive Evaluation of Algorithms and Training Paradigms: We benchmark eight continual learning algorithms across four major categories and systematically compare reinforcement learning with supervised fine-tuning paradigms; (3) Carefully Curated Tasks: We select and organize 16 datasets from existing work, covering six challenging domains. Our key findings include: (i) Models with stronger general capabilities exhibit greater robustness to forgetting during continual learning; (ii) Reasoning chains degrade more slowly than final answers, supporting the hierarchical forgetting hypothesis; (iii) The effectiveness of continual learning algorithms is highly dependent on both model capability and task order; (iv) In reinforcement learning settings, incorporating KL-divergence constraints helps maintain policy stability and plays a crucial role in mitigating forgetting. MLLM-CTBench establishes a rigorous standard for continual instruction tuning of MLLMs and offers practical guidance for algorithm design and evaluation.