Reinforcement learning with verifiable reward (RLVR) has been shown to enhance reasoning in large language models, yet its effectiveness on long‑horizon multimodal tasks remains under‑examined. Visual Sudoku puzzles represent a prototypical challenge, requiring multi‑step deductive reasoning that intertwines local constraints with global dependency chains. To address this, we introduce the FG‑Sudoku benchmark, which categorizes puzzles by grid size and the number of required deduction steps. Evaluations reveal that state‑of‑the‑art multimodal LLMs fail on complex long‑range dependencies such as 9×9 puzzles, resulting in near‑zero success rates during fine‑tuning and severe reward sparsity that stalls standard RLVR methods. We therefore propose Adaptive Curriculum Group Relative Policy Optimization (AdaC‑GRPO). Unlike rigid or prior‑based curricula, AdaC‑GRPO dynamically assembles training batches by tracking the model’s Zone of Proximal Development via online rollout accuracy, thereby maintaining effective learning despite sparse rewards. Experiments demonstrate that, after applying AdaC‑GRPO, Qwen2.5‑VL‑7B approaches the performance of Gemini‑2.5‑Flash and bridges 62.2% of the overall accuracy gap to Gemini‑2.5‑Pro; its generalizability is further confirmed on InternVL‑3.5 and Qwen3‑VL.
Review: AdaC‑GRPO’s ability to adaptively gauge model competence and adjust task difficulty on‑the‑fly mitigates reward sparsity in long‑horizon visual Sudoku, offering a practical curriculum learning framework for multimodal reasoning tasks.