cauchyfanupdate/gensis is a public research code repository that provides data generation for VLM visual reasoning, benchmark evaluation, and reinforcement learning training code. The README states that the related paper has been accepted for publication and that the implementation language is Python. The repository presents no separate recent release information.
Genesis's data construction method creates visual edits that change the correct answer, then diversifies other visual attributes while keeping the task and adding reasoning annotations. The repository covers this data generation process, the construction and evaluation of the UNREAL benchmark, and reinforcement learning training that uses reward criteria. Reinforcement learning is a way of training a model by assigning rewards to outcomes. Reward learning is tied to criteria that evaluate intermediate reasoning, and thus deals with visual reasoning. The README includes a table of evaluation scores for the baseline model Qwen3-VL and the Genesis model, and explains that these values are results reported in the paper. It also provides guidance on data generation, evaluation, and training, along with links to models and datasets.
That said, the scores are values reported in the paper, so their separate verification cannot be confirmed from the repository materials alone. No license is stated either. The code connecting generation, evaluation, and training is useful for examining the research procedure, but it is difficult to judge the reproducibility of the results from publicly available information alone.