AgenticGen: Reward-Guided Agentic Video Generation for Advertising

Details

TikTok researchers introduce AgenticGen, a framework that decomposes advertising video generation into two trainable reasoning stages: strategy selection (how to present a product) and draft generation (what to actually produce). By making these stages explicit, the system creates optimization targets that live ad performance metrics can directly supervise. AgenticGen learns two complementary reward signals: a performance-based reward derived from accumulated CTR, CVR, and conversion data, and a rubric-based reward aligned with human quality standards. Training proceeds in two stages — DPO first aligns agentic policies toward online preferences derived from live feedback, followed by GRPO (process and outcome reward) to refine both stages further. Offline experiments validate the reward models. Live A/B tests in the production TikTok advertising system show CTR +2.72%, CVR +2.63%, and Advv (advertising value) +9.61% over a supervised fine-tuning baseline. The 9.61% Advv improvement is the headline number — at TikTok's scale, even percentage-point gains translate to substantial revenue. This is one of the few published examples of GRPO applied to video generation in a production ad system, extending the technique beyond language tasks into multimodal content optimization.