AgenticGen: Reward-Guided Agentic Video Generation for Advertising
AgenticGen ties ad-video generation to online feedback, then reports gains inside TikTok’s ad system.
The paper frames advertising video generation as a product-conditioned reasoning task, not just clip synthesis. Its framework splits the process into strategy selection and draft generation, with rewards learned from online performance feedback and human-quality rubrics. The authors say DPO and GRPO optimization improved CTR by 2.72%, CVR by 2.63%, and Advv by 9.61% over an SFT baseline in online A/B tests. HF Daily Papers' note
The paper frames advertising video generation as a product-conditioned reasoning task, not just clip synthesis. Its framework splits the process into strategy selection and draft generation, with rewards learned from online performance feedback and human-quality rubrics. The authors say DPO and GRPO optimization improved CTR by 2.72%, CVR by 2.63%, and Advv by 9.61% over an SFT baseline in online A/B tests. HF Daily Papers' note
score 4