Tool-using wrapper lifts video model scores by 19 points on its authors' benchmark
single source· 1 articles · confidence: medium · first seen 2026-09-20 20:00 UTC
What this means for you
Nothing to build on yet: this is a paper, with no code, weights, price or release date given, and the headline number comes from a benchmark its own authors built, so it is not comparable with published video scores. The transferable finding is the shape — the trained tool-use layer kept working when the generator underneath it improved.
VideoGen-Agent wraps a text-to-video model and calls external tools in a loop rather than generating from the prompt alone, using what it observes to pick the next call. The authors fine-tuned it on teacher trajectories, then refined it with reinforcement learning against a reward for tool-call validity and video quality. They also built VABench, 600 held-out prompts covering procedural knowledge, identity preservation, physical consistency, scene composition and multi-shot structure. On it the agent scores 75.6, up 19.1 points from its base generator's 56.5. A better generation tool lifts that to 86.1 with no further agent training; human raters preferred it to the strongest standalone baseline in 84.3% of comparisons.
Key facts
- ·VideoGen-Agent scores 75.6 on VABench against 56.5 for its base text-to-video generator, a 19.1-point improvement. source
- ·VABench is a held-out benchmark of 600 prompts introduced with the paper, covering procedural knowledge, single- and multi-entity identity preservation, physical consistency, scene composition and multi-shot temporal structure. source
- ·Swapping in stronger generation tools raises the VABench score to 86.1 with no additional agent training. source
- ·Human raters preferred the upgraded configuration over the strongest standalone baseline in 84.3% of comparisons. source
- ·The agent was trained by supervised fine-tuning on teacher-generated trajectories, then refined with reinforcement learning against a category-aware hybrid reward. source
- ·The paper reports no evaluation date for the VABench scores. source
What the sources say
- Hugging Face Daily Papers (research) — Describes an agent that calls augmentation, generation and verification tools, plus the benchmark it was scored on.
Sources
The original reporting. Follow these — they did the work.
- Hugging Face Daily PapersVideoGen-Agent: Reinforcing Video Generation Agents2026-09-20