SyncLoop: A Multimodal Dual-Loop Framework for Self-Improving Mathematical Reasoning

Recent advances in multimodal large language models (MLLMs)have shown impressive reasoning capabilities. However, further enhanc-ing existing MLLMs necessitates high-quality vision-language datasetswith carefully curated task complexities, which are both costly and chal-lenging to scale. Although recent self-improving models that iterativelyrefine themselves offer a feasible solution, they still suffer from two corechallenges: (i) most existing methods augment visual or textual data sep-arately, resulting in discrepancies in data complexity (e.g., over-simplifieddiagrams paired with redundant textual descriptions); and (ii) the evo-lution of data and models is also separated, leading to scenarios wheremodels are exposed to tasks with mismatched difficulty levels. To ad-dress these issues, we propose SyncLoop, an automatic, closed-loopself-improving framework that jointly evolves both training data andmodel capabilities. Specifically, given a base dataset and a base model,SyncLoop enhances them by a cross-modal data evolution loop and adata-model evolution loop. The former loop expands the base datasetby generating complex multimodal problems that combine structuredtextual sub-problems with iteratively specified geometric diagrams ormathematical functions, while the latter loop adaptively selects the gen-erated problems based on the performance of the base model, to conductsupervised fine-tuning and reinforcement learning alternately. Conse-quently, our method continuously refines its model and training data,and consistently obtains considerable performance gains across multiplemathematical reasoning benchmarks.