The authors frame visual Sudoku as a testbed for long-horizon multimodal reasoning, since solving N by N grids requires chaining deductive steps across both local constraints and global dependency chains.
Adaptive curriculum training lets vision-language models solve hard visual Sudoku puzzles
A competence-driven reinforcement learning method samples tasks in the model's Zone of Proximal Development, lifting hardest-tier accuracy from near zero to 0.219.
Chinese Tech
Sida Li · Yuanxin Liu · Linli Yao · Shicheng Li · Xu Sun · Haoli Bai · +1 more
Peking University · Huawei Technologies Co., Ltd.
Research Digest··3 min read
The authors introduce FG-Sudoku, a graded benchmark of visual Sudoku puzzles, and AdaC-GRPO, a curriculum reinforcement learning method that adaptively samples tasks within the model's Zone of Proximal Development.
Why this paper
From Huawei Technologies Co., Ltd. and Peking University
In one line
Adaptive curriculum learning via online rollout accuracy enables RLVR to solve long-horizon visual Sudoku puzzles where standard methods fail.
What we could check
- ·No code link found
- ·No weights link found
- ·No dataset link found
- ·No compute details found
- ·No stated limitations found
- ✓Reports numbers on named benchmarks
Observed from the paper text and links we have. Absence here means we did not find it, not that it does not exist.
§