APT: Feedback-Driven Agentic Post-Training for Text-to-Image Generation
Sirui Zhang, Fangyun Wei, Dong Chen, Jinjing Zhao, Yitong Wang, Yunuo Chen, Baining Guo, Zheng-Jun Zha
University of Science and Technology of China · Microsoft Research Asia · University of Sydney · Fudan University · Shanghai Jiao Tong University
Published October 13, 2026 · on the feed October 15, 2026 · ▲ 22 · arXiv 2610.18097
The abstract
Post-training is crucial for improving text-to-image models, yet existing pipelines typically rely on manually curated training data and fixed training distributions. We introduce APT (Agentic Post-Training), a fully automated framework that adaptively post-trains a base model according to its evolving capability profile. APT coordinates three agents to construct evaluation sets from user requirements, diagnose model-specific strengths and failure modes, and generate reinforcement learning data with an adaptive prompt distribution that emphasizes underperforming capabilities. By repeatedly evaluating, diagnosing, and updating the generator, APT lets the current model's behavior shape its next training tasks, forming a feedback loop inspired by recursive self-improvement. For general-purpose settings without user-specified evaluation scenarios, we further develop an evaluation harness that enables broad and diverse evaluation set generation. Experiments across multiple base models, benchmarks, and generation architectures demonstrate that APT is model-agnostic, benchmark-agnostic, architecture-agnostic, and fully automated.
The Brief
2:06A short podcast summary of the paper. Written and voiced by AI.
Transcript
APT. Feedback-Driven Agentic Post-Training for Text-to-Image Generation. By Sirui Zhang and seven others, from the University of Science and Technology of China, Microsoft Research Asia, and three other institutions. Published on the thirteenth of October. Its code repository is a placeholder: a readme, and nothing to run yet.
Post-training matters a great deal for text-to-image models. Yet the usual pipelines lean on training data curated by hand, and on a training distribution that is fixed in advance. Neither one responds to what the particular model in front of them is good or bad at. Once the data is chosen, training proceeds the same way whether the model has already mastered a skill or is still failing at it.
APT, for Agentic Post-Training, is a fully automated framework that post-trains a base model adaptively, according to its capability profile as that profile evolves. The current model's behavior shapes its next training tasks, a feedback loop the authors describe as inspired by recursive self-improvement.
Three agents divide the work. One constructs evaluation sets from user requirements. One diagnoses the model's specific strengths and failure modes. One generates reinforcement learning data, drawing prompts from an adaptive distribution that emphasizes the capabilities where the model underperforms. The cycle of evaluating, diagnosing and updating the generator then repeats. For general-purpose settings where the user has not specified evaluation scenarios, the authors also build an evaluation harness that generates broad and diverse evaluation sets. Each pass through the loop starts from the model as it now is, so the emphasis moves as old weaknesses close and new ones appear.
Experiments run across multiple base models, multiple benchmarks, and different generation architectures. The claim they support is that APT is model-agnostic, benchmark-agnostic, architecture-agnostic, and fully automated.
What changes is who decides what to train on next. Instead of a person assembling data once, the model's own weaknesses, measured fresh each round, set the curriculum.
More with an account
Listen to every day’s papers
Daily playlists, deep dives, quests for the papers, and quizzes on their key terms.
Join free