huangrh99/alphagrpo

AlphaGRPO is the official repository accompanying a research paper accepted at ICML 2026, a top machine learning conference. The paper introduces a training method for unified multimodal AI models, meaning models that can both understand and generate across text and images within a single system, such as the BAGEL model it references as an example. The core idea is a reinforcement learning approach that teaches these models to reflect on and refine their own image generation, using what the paper calls a decompositional verifiable reward. In plain terms, instead of judging a generated image with one overall score, the training process breaks the evaluation into smaller, checkable pieces of feedback, which helps guide the model toward better results, including generation that involves reasoning about the prompt and self correction after an initial attempt. The codebase supports several different reinforcement learning methods for image generation, named FlowGRPO, DiffusionNFT, and AWM, as well as a method called GRPO for text generation. According to the README, the paper reports improved results on text to image generation benchmarks and on an image editing benchmark called GEdit-Bench, even though the method was not specifically trained for editing tasks. As of this README, only the paper itself has been released on arXiv. The actual training code and trained model weights are still going through internal review and have not been published yet, so this repository currently serves mainly as a landing point for the paper, its results, and citation information rather than a runnable tool. It is released under the Apache License 2.0 in preparation for that future code release.