WeVisCut: The Whole Editing Room in One Agent From Narration to Long-Form Video

WeVisCut connects a stateful script-to-video agent harness, a four-stage pipeline that recovers executable supervision from finished videos, and a single 9B policy trained through SFT, runtime-verified RFT, Guided-OPD, and RCF-DAPO, evaluated on a 100-task benchmark with five metrics.

Existing agentic editors often coordinate long-form script-to-video production through multi-stage workflows driven by general models, leaving unclear how to train a compact task-adapted policy that balances quality and token cost. Finished videos omit the tool calls, intermediate artifacts, and failures needed for executable supervision. The challenge is therefore to connect stateful execution, executable supervision, and policy learning.

The WeVisCut harness maintains shared production state and uses one policy across orchestrator, generator, and critic roles, with typed tool contracts, runtime validation, blockers, inspection, bounded repair, and verified delivery. A four-stage data pipeline performs editing inversion, verified replay, observe–reflect–repair, and trajectory materialization on 3,750 finished video projects, yielding 134,771 role-specific examples. The shared Qwen3.5-9B policy is first trained with unified SFT and runtime-verified rejection-sampled fine-tuning. Phase 2 applies Guided-OPD with teacher intervention probability beta_n reduced to zero, so training ends with student-controlled executions. Phase 3 applies RCF-DAPO, which restricts reward comparisons to matched roles and production contexts.

• In the shared tool environment, WeVisCut improves Overall Score over its Qwen3.5-9B initialization with gains in all five metrics; SCF and VNSM increase by 18.6 and 15.4 points respectively. §6.2
• WeVisCut reaches an Overall Score close to Claude Opus 5, with higher means on four metrics and slightly lower VNSM, while using far fewer generated tokens per successful output minute than Claude Opus 5 and fewer than its base policy. §6.2
• In the native-workflow comparison on the same 100 tasks, WeVisCut has the highest mean score on all five metrics and lower wall-clock and estimated monetary cost per successful output minute than Crayotter and VideoAgent. §6.3
• Cumulative training stages raise Overall Score from the base policy through SFT, RFT, Guided-OPD, and RCF-DAPO; SFT alone raises ERY from 70.4 to 87.6. §6.4
• Blinded human evaluation gives WeVisCut the highest mean rating on all six dimensions, with overall preference 4.56 versus 3.48 for Crayotter and 3.46 for VideoAgent; the largest gains are in narrative coherence. §6.5

• The paper states that WeVisCut is limited by the current HyperFrames template library, so uncommon visual languages may be approximated using the closest available structure, motion, and style.
• The paper states that the dataset and project-disjoint split cannot cover all domains, editing conventions, runtimes, or asset distributions, and verified replay guarantees executability only within the registered runtime.
• The paper states that long-form video quality depends on the stability of the video retriever, particularly for instruction-based retrieval.
• The paper states that VNSM relies on a VLM, and evidence isolation, frozen anchors, agreement reporting, and human calibration cannot eliminate judge bias or service drift.

• A single policy with shared parameters serves orchestrator, generator, and critic roles, and each call conditions on the task, role, production state, history, and tool schemas. §3.1
• The runtime checks argument schemas, artifact references, and execution prerequisites; only validated results are committed, while rejected calls return typed blockers. §3.2
• Quality inspection combines structured checks with a VLM-Judge over spotlight renders, and bounded repair preserves still-valid artifacts, so correcting a decoration need not regenerate unchanged narration audio. §3.2
• Executable supervision is built through editing inversion, verified replay, observe–reflect–repair, and trajectory materialization. §4.1
• Role-specific examples supervise tool selection and arguments, generated content, and critic decisions; loss applies only to model-authored tokens, and the split is by source project. §4.2
• Guided-OPD selects teacher outputs with probability beta_n and student outputs otherwise, trains the student on both branches, and reduces beta_n to zero so training ends with student-controlled executions. §5.2
• RCF-DAPO compares candidate outputs only within the same role and matched production context, using role-specific rewards and DAPO's Clip-Higher and soft overlong shaping. §5.3
• The benchmark contains 100 held-out script-to-video tasks; five metrics lie in [0,100], Overall Score is their equal-weight mean, and Production Cost Efficiency is reported separately. §6.1