Notice: The reproducibility variables underlying each score are classified using an automated LLM-based pipeline, validated against a manually labeled dataset. LLM-based classification introduces uncertainty and potential bias; scores should be interpreted as estimates. Full accuracy metrics and methodology are described in Coakley et alK. L. Coakley, T. Snelleman, H. Hoos, and O. E. Gundersen, "The embrace of open science: An analysis of a decade of AI research and 56 800 conference papers," Under Review, 2026..

Reinforcement Learning Meets Masked Generative Models: Mask-GRPO for Text-to-Image Generation

Authors: Yifu Luo, Xinhao Hu, Keyu Fan, Haoyuan Sun, Zeyu Chen, Bo Xia, Tiantian Zhang, Yongzhe Chang, Xueqian Wang

NeurIPS 2025 | Venue PDF | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental 4 Experiments
Researcher Affiliation Academia 1Tsinghua University 2Technical University of Munich
Pseudocode No No explicit pseudocode or algorithm blocks are provided in the paper. The methods are described using prose and mathematical formulations.
Open Source Code Yes The code is available on https://github.com/ xingzhejun/Mask-GRPO.
Open Datasets Yes We conduct our experiments using the training set of LAION dataset [53] without accompanying images.
Dataset Splits Yes We conduct our experiments using the training set of LAION dataset [53] without accompanying images. Following Stable Diffusion [39, 5], we evaluate MASK-GRPO with standard T2I benchmarks Gen Eval [55] and FID [56] on the MSCOCO dataset [57].
Hardware Specification Yes All experiments are conducted on 16 NVIDIA A100 80GB GPUs.
Software Dependencies No The paper mentions using CLIP and Image Reward as reward models and Adam as an optimizer, but does not provide specific version numbers for software libraries or environments like Python, PyTorch, or CUDA.
Experiment Setup Yes The learning rate is set as 3e 6 and the group size G set as is 6. We utilize Adam as our optimizer and set beta to 0.95, and we set the batch size to 96 (6 rollouts per GPU for one prompt).