AgenticGen: Reward-Guided Agentic Video Generation for Advertising

AgenticGen. Reward-Guided Agentic Video Generation for Advertising. By Xingyuan Bu and eleven others, from ByteDance, Tsinghua University, and one other institution. Published on the thirty-first of August. No code repository is linked to it.

Generating an advertising video is more than video synthesis. It is a reasoning problem conditioned on a product, and success is measured by online business metrics. Recent video foundation models can generate realistic clips from multimodal inputs, but they do not optimize how a product should be turned into an effective advertisement. Nor do they learn, from online business feedback, how to generate better next time. A clip can look entirely realistic and still be a poor advertisement, and nothing in a model trained only for realism would notice. The feedback that would tell it, how the advert actually performs once it runs, arrives after generation and has no route back in.

AgenticGen closes that loop with a reward-guided agentic framework. It splits advertising video generation into two trainable reasoning stages: strategy selection, then draft generation. Splitting it this way exposes targets that online business feedback can supervise directly. A single end-to-end generator offers no obvious place to apply a business signal. Two explicit stages, one choosing the approach and one producing the draft, each give that signal something concrete to train.

Two rewards guide the training. One is performance-based, learned from online feedback accumulated over time. The other is rubric-based and complements the first, aligned with human standards of quality. The pairing matters because each covers what the other can miss: one tracks what audiences respond to, the other what people judge to be good work. Both are used to supervise policy optimization, which proceeds in two steps. DPO first moves the agentic policies toward online preferences. GRPO then refines both stages further, using rewards for the process as well as for the outcome.

Offline experiments validate the reward models and the successive rounds of policy optimization. The decisive test is online, in A B experiments inside the TikTok advertising system. Against the SFT baseline, AgenticGen after DPO and GRPO improves CTR by two point seven two percent and CVR by two point six three percent. Advv improves by nine point six one percent.

Much of the contribution lies in the framing. By treating an advert as a strategy decision followed by a draft, and by learning rewards from what actually performs, the generator is trained against the business outcome it exists to serve rather than against realism alone.