JustLLMGRPO: LLM Prompt Optimization Improves Chest X-Ray Generation

A new arXiv preprint (2608.08046) introduces JustLLMGRPO, a method that applies Group Relative Policy Optimization (GRPO) to the language model (LLM) prompt policy while keeping the Sana image generator frozen. The research demonstrates that reformulating text prompts with an unmodified LLM can significantly improve the quality of text-conditioned chest X-ray (CXR) generation. Specifically, using a CXR-adapted Sana generator, one-pass reformulation reduced RadDINO-FID from 54.225 to 27.572, indicating better fidelity to the source prompts. However, unconstrained reformulation reduced BioViL-T alignment with source prompts from 0.695 to 0.609, showing a trade-off between image quality and semantic alignment. To address this, JustLLMGRPO applies GRPO to the LLM prompt policy, optimizing the reformulation to balance these objectives. The work highlights that the generator-centric view of existing methods leaves an underexplored optimization dimension in the prompt space. The study is relevant to the intersection of AI, medical imaging, and natural language processing, and was published on arXiv under the title 'JustLLMGRPO: Radiographic Control for Chest X-Ray Generation'.