PKU, Kling AI Reviso Lifts Qwen3.5-9B F1 by 15.4 on MCTS-VCB

Reviso, an agentic video captioner from Peking University, Kling AI Research and six other universities, lifts its Qwen3.5-9B backbone by 15.4 F1 points on the MCTS-VCB benchmark by planning what to look at before it looks at frames. The paper, posted on Hugging Face, argues that detailed video captions miss the point when models weight every visual equally: "not all visual evidence deserves equal attention, as the importance of visual information depends on the content of each video."

From a low-frame-rate global preview, Reviso builds a hierarchical Context-Focus-Points plan that names the setting, prioritises aspects and spells out concrete observation targets with the tools to resolve them, including image cropping and OCR. It then revisits those segments at higher frame rate, interprets what it finds, and updates the plan.

Training uses 10,000 trajectories for supervised fine-tuning and 20,000 for reinforcement learning, drawn from LLaVA-Video-178K, Vript, Ego4D and YouTube. The team runs LoRA SFT for two epochs followed by 200 steps of GRPO with three reward components: a Key-Fact Reward, a Plan Alignment Reward, and an Unproductive Tool-Call Penalty that "adjusts observation-cost penalties based on caption quality."

With every model capped at 128 input frames, Reviso hits F1 of 39.5 on DREAM-1K and 60.9 on MCTS-VCB, beats evaluated open-source baselines across four captioning benchmarks, and lifts downstream video-QA accuracy on LVBench by 4.9 points over the Qwen3.5-9B backbone. The authors note that among evaluated agentic workflows on DREAM-1K, "only GPT-6 Astra paired with Codex achieving a higher F1 score" than Reviso. Code is up at github.com/pkucs-Ltf/Reviso, landing alongside a run of agentic-harness papers we have been tracking in open-source AI this week.