Adaptive learning boosts solving visual sudoku puzzles in ai models
AdaC-GRPO: Towards Mastering Long-Horizon Visual Sudoku Puzzles via Adaptive Curriculum Learning
Artificial Intelligence
Summary
Solving visual Sudoku puzzles is hard for AI models because they require many detailed steps and understanding of the whole puzzle at once. The authors created a new Sudoku benchmark to test these challenges and found that current AI models struggle, especially with large puzzles. They proposed a new training method called AdaC-GRPO, which adapts the difficulty of problems based on the model’s progress, helping it learn better over time. This approach significantly improved the AI’s puzzle-solving skills compared to previous methods.
What this means in practice
- •For multimodal ai developers: Improve training methods for AI models that analyze and reason with images and text by dynamically adjusting problem difficulty to their skill level.
- •For game ai engineers: Design AI that can solve complex, long-step puzzle games by using adaptive curricula that track model learning progress.
Authors
Sida Li, Yuanxin Liu, Linli Yao, Shicheng Li, Xu Sun, Haoli Bai, Xiaohui Li
Abstract
While reinforcement learning with verifiable reward (RLVR) enhances reasoning in large language models, its efficacy in long-horizon multimodal tasks remains underexplored. Visual Sudoku puzzles pose a representative challenge, requiring multi-step deductive reasoning over both local constraints and global dependency chains. To this end, we introduce FG-Sudoku, a fine-grained benchmark categorized by grid size and required deductive steps. Evaluations reveal that frontier MLLMs fail on complex long-range dependencies (e.g., 9*9 puzzles). During post-training, this near-zero success rate causes severe reward sparsity, halting standard RLVR. We propose Adaptive Curriculum Group Relative Policy Optimization (AdaC-GRPO) to overcome this. Unlike rigid or prior-based schedules, AdaC-GRPO dynamically constructs training batches by tracking the model's "Zone of Proximal Development" via online rollout accuracy. Experiments demonstrate that AdaC-GRPO enables Qwen2.5-VL-7B to approach the performance of Gemini-2.5-Flash while bridging 62.2% of the overall accuracy gap to Gemini-2.5-Pro, with its generalizability further validated across InternVL-3.5 and Qwen3-VL.