Researchers have developed a framework for conversational systems to provide follow-up edit suggestions in image-creation tasks. The system uses real online data and user feedback to optimize the policy through reinforcement learning. The framework significantly outperforms baselines on both automatic and human evaluations, reducing visual inconsistency by 75% and improving recommendation metrics such as click-through rate and conversation turns per user. --- Why it matters: This matters because conversational assistants increasingly need to handle image-creation tasks, which require multimodal recommendations that reflect user preferences. This framework addresses the challenge of providing useful follow-up edit suggestions in these tasks. Source: https://arxiv.org/abs/2608.07565