aiqbench analysis
This paper addresses the critical challenge of long-horizon reasoning in multimodal large language models by introducing AdaC-GRPO, an adaptive curriculum learning method that mitigates reward sparsity during reinforcement learning with verifiable rewards. By dynamically adjusting training batches based on a model's "Zone of Proximal Development," the approach significantly improves performance on complex visual Sudoku tasks, narrowing the gap between open-source and proprietary models. The work highlights essential advancements in algorithmic efficiency for solving multi-step deductive problems without relying on external priors.
AdaC-GRPO: Towards Mastering Long-Horizon Visual Sudoku Puzzles via Adaptive Curriculum Learning
arXiv:2610.13184v1 Announce Type: new 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.
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.
Artificial Intelligence (cs.AI)
arXiv:2610.13184 [cs.AI]
(or arXiv:2610.13184v1 [cs.AI] for this version)
https://doi.org/10.48550/arXiv.2610.13184 Focus to learn more arXiv-issued DOI via DataCite (pending registration)
arXivLabs is a framework that allows collaborators to develop and share new arXiv features directly on our website.
Both individuals and organizations that work with arXivLabs have embraced and accepted our values of openness, community, excellence, and user data privacy. arXiv is committed to these values and only works with partners that adhere to them.
Have an idea for a project that will add value for arXiv's community? Learn more about arXivLabs.
Full text under the publisher's open licence · 242 words
Related coverage
[AINews] DeepSeek v4.1-Flash: 763B-P8B-D16B novel causal Encoder–Decoder architecture with vision marks the Return of the Whale
Latent Space · Sep 12, 2026
Daniel Ritter
file770.com · Aug 29, 2026
Dyna-2: General Purpose Physical Agents with Hour-Long Autonomy — Jason Ma, Dyna Robotics
Latent Space · Oct 13, 2026
SkillWeave: Weaving Heterogeneous Demonstrations into Long-Horizon Manipulation Skills
arXiv · Oct 9, 2026