arXivDaily arXiv每日学术速递 周一至周五更新
arXiv周末暂无论文更新,休息一下吧,周末愉快~~
arXiv 2610.18097cs.CVcs.AI
收藏

APT:反馈驱动的智能体式文本到图像生成模型后训练

APT: Feedback-Driven Agentic Post-Training for Text-to-Image Generation

University of Science and Technology of China(中国科学技术大学);Microsoft Research Asia(微软亚洲研究院);University of Sydney(悉尼大学);Fudan University(复旦大学);Shanghai Jiao Tong University(上海交通大学)

Sirui Zhang, Fangyun Wei, Dong Chen, Jinjing Zhao, Yitong Wang, Yunuo Chen, Baining Guo, Zheng-Jun Zha

AI总结:

APT是一种完全自动化的智能体式后训练框架,通过协调三类智能体形成反馈循环,可自适应提升文本到图像生成模型的性能,且具备模型、基准、架构无关的特性。

摘要:

后训练对于提升文本到图像模型至关重要,但现有流程通常依赖人工整理的训练数据和固定的训练分布。我们提出APT(Agentic Post-Training,智能体式后训练),这是一种完全自动化的框架,可根据基础模型不断变化的能力特征对其进行自适应后训练。APT协调三个智能体,根据用户需求构建评估集,诊断模型特定的优势和失败模式,并生成强化学习数据,其自适应提示分布会重点关注表现不佳的能力。通过反复评估、诊断和更新生成器,APT让当前模型的行为决定其下一个训练任务,形成受递归自我改进启发的反馈循环。对于没有用户指定评估场景的通用设置,我们进一步开发了评估工具,可生成广泛多样的评估集。在多个基础模型、基准和生成架构上开展的实验表明,APT具有模型无关、基准无关、架构无关且完全自动化的特性。

英文摘要:

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.

机构、总结、中文摘要均由AI分析而来,可能存在错误,请以论文正文为准。

↑