Abstract

While current Multimodal Large Language Models (MLLMs) have demonstratedproficiency in reasoning tasks such as mathematics and logic, their capacityfor long-chain reflective reasoning, a prerequisite for solving complexreal-world problems, remains largely underexplored. In this work, we firstconduct an extensive empirical investigation to evaluate this capability.Leveraging a carefully designed data synthesis engine, we construct MM-HELIX, amultimodal benchmark consisting 1,260 samples of 42 challenging synthetic tasksthat require iterative thinking and backtracking. Empirical results on thisbenchmark reveal that existing MLLMs exhibit significant performance deficitsin long-chain reflective reasoning. To address this limitation, we generatepost-training data and further explore learning paradigms for exploiting suchdata. We first develop the Step-Elicited Response Generation pipeline to createMM-HELIX-100K, a large-scale dataset of 100k high-quality, reflective reasoningtraces for instruction-tuning stage. Given that standard Reinforcement Learningfails on complex tasks due to sparse reward signals and catastrophic forgettingafter Supervised Fine-Tuning, we propose Adaptive Hybrid Policy Optimization(AHPO), a novel training strategy that dynamically unifies offline supervisionand online optimization into a single stage. This strategy enables the model tolearn from expert data when rewards are sparse and conduct independentexploration once proficient. When applied to the Qwen2.5-VL-7B baseline, ourmethod achieves a +18.6\% accuracy improvement on MM-HELIX benchmark anddemonstrates strong generalization with a +5.7\% average performance gain ongeneral mathematic and logic tasks. Our work demonstrate that reflectivereasoning in MLLMs can be effectively learned and generalized, paving the wayfor developing more capable MLLMs.