Advertising video generation is more than a synthesis task; it is a product‑conditioned reasoning problem whose success is measured by online business metrics. Existing video foundation models can produce realistic clips from multimodal inputs, yet they do not optimize how a product should be turned into an effective ad nor leverage online feedback to improve future generations. To close this loop we introduce AgenticGen, a reward‑guided agentic framework that decomposes the task into two trainable reasoning stages—strategy selection and draft generation—thereby exposing optimization targets that can be supervised by online business feedback.
AgenticGen learns a performance‑based reward from accumulated online signals and a complementary rubric‑based reward aligned with human quality standards. Policy optimization proceeds in two steps: DPO (Direct Preference Optimization) moves the agentic policies toward online preferences, and GRPO (Generalized Reward‑based Policy Optimization) further refines both stages using process and outcome rewards.
Offline experiments confirm the validity of the reward models and the benefits of successive policy optimization. In online A/B tests on TikTok’s advertising platform, AgenticGen after DPO and GRPO improves CTR by 2.72%, CVR by 2.63%, and advertising revenue (Advv) by 9.61% over the SFT baseline.
Review