VideoGen-Agent: Reinforcing Video Generation Agents
A tool-using video agent beat its base text-to-video generator by 19.1 points on the authors’ held-out benchmark.
VideoGen-Agent is trained to coordinate augmentation, generation, and verification tools through multi-turn decisions. The paper says it targets prompts that need specialized knowledge, identity preservation, physical consistency, ordered events, scene composition, and multi-shot temporal structure. On VABench, its score rose from 56.5 to 75.6 over the base generator, and to 86.1 when the generation tools were upgraded without retraining the agent. Human raters preferred that upgraded setup over the strongest standalone baseline in 84.3% of comparisons. HF Daily Papers' note
VideoGen-Agent is trained to coordinate augmentation, generation, and verification tools through multi-turn decisions. The paper says it targets prompts that need specialized knowledge, identity preservation, physical consistency, ordered events, scene composition, and multi-shot temporal structure. On VABench, its score rose from 56.5 to 75.6 over the base generator, and to 86.1 when the generation tools were upgraded without retraining the agent. Human raters preferred that upgraded setup over the strongest standalone baseline in 84.3% of comparisons. HF Daily Papers' note
score 5