The hard part of long-form video editing was never gluing a few clips together. It was keeping a pile of interdependent decisions coherent across dozens of shots: change one shot's duration and every later timestamp shifts with it; when the retrieved footage is wrong, swapping the asset is not enough, the shot plan has to be revisited too. WeVisCut takes a different route than orchestrating a workflow with general-purpose models. It puts the whole editing room inside a single 9B policy that decides for itself which tool to call next. Full paper
Mining supervision out of finished videos
The real bottleneck is data. A finished video keeps the result but discards the tool calls, intermediate artifacts, and failures, and agent training needs exactly those. WeVisCut builds supervision with a four-stage data engine. Editing inversion first recovers the script-to-shot intent from a finished video. Verified replay then executes that intent with production tools. When replay surfaces defects, observe-reflect-repair applies bounded corrections while preserving artifacts that already succeeded. Trajectory materialization finally organizes the recorded executions into training examples carrying tool feedback and repair outcomes.
The paper reports 134,771 training examples constructed from 3,750 completed video projects. The official repository ships the collection pipeline as data_construct_pipeline, and states that the released runtime serves the policy the same 12 tools and the same state JSON it saw during training.
Three phases take the score from 74.5 to 90.5
Training runs in three phases. Phase 1 is unified SFT plus runtime-verified rejection-sampled fine-tuning, which teaches tool use and recovery from failure. Phase 2 applies Guided-OPD to learn recovery on states the student itself reached, with a Qwen3.5-397B-A17B FP8 teacher whose intervention probability decays from 0.8 to 0 over 70 steps. Phase 3 introduces RCF-DAPO, which restricts DAPO's reward comparisons to the same role and the same production context.
The ablation table spells out the path. The Qwen3.5-9B base scores 74.5 overall, SFT lifts it to 79.5, RFT brings it to 83.5, Guided-OPD to 84.9, and RCF-DAPO to 90.5. The last two phases contribute more than the earlier ones.
The cost ledger, and one metric it loses
In a shared tool environment the paper reports overall scores of 90.5 for WeVisCut, 89.9 for Claude Opus 5, 84.6 for Gemini 3.1 Pro, 81.5 for Qwen3.5-397B-A17B, 77.7 for Qwen3.6-35B-A3B, and 75.9 for Qwen3.6-27B. The quality gap is genuinely narrow. The cost column is not. Per successfully delivered output minute, WeVisCut spends 2,377 generated tokens against 78,763 for Opus 5, roughly 3% of the total, which the paper describes as nearly 33x fewer tokens. Calls per minute run 3.1 against 11.6.
Against systems running their own native workflows, WeVisCut takes 2.1 minutes and 0.09 USD per output minute, while Crayotter needs 8.6 minutes and 0.97 USD, and VideoAgent 10.4 minutes and 0.58 USD.
One metric goes the other way. On VNSM, narrative consistency, WeVisCut scores 74.6 against 74.8 for Opus 5, and VNSM is the only one of the five metrics scored by a vision-language model.
Human ratings go higher, and the paper brakes itself
Three qualified annotators rated anonymized videos on the same 100 held-out tasks. The paper reports WeVisCut's mean above the other two systems on all six dimensions, with an overall preference of 4.56 against 3.48 for Crayotter and 3.46 for VideoAgent. The clearest gaps appear on narrative coherence, at 1.23 and 1.28 points.
The paper does not oversell this. The text states that these descriptive means do not establish statistical significance and do not validate individual automatic metrics. The limitations section is equally candid: visual style is bounded by the current HyperFrames template library, verified replay only guarantees executability inside the registered runtime, and VNSM leans on a vision-language model, where evidence isolation, frozen anchors, and human calibration still cannot eliminate judge bias or service drift.
My read
Reading this as "a 9B beats Opus 5" is not accurate. That 0.6-point gap between 90.5 and 89.9 sits inside the range the paper itself says carries no statistical significance. The hard number is token cost: nearly 33x, which means the bill for the same amount of finished footage can differ by an order of magnitude.
The data engine is the part worth copying. Editing inversion plus verified replay is a pipeline that is indifferent to model and modality, so any team building long-video, long-document, or long-horizon workflow agents can reuse it. The architectural choice is worth noticing too: 12 tools, one tool call per step, guardrails that reject illegal calls instead of substituting a decision of their own, and stability bought with many cheap calls.
The lost VNSM metric points at something a small task-specific model cannot fix: when "is this good" has to be judged by another vision-language model, the 9B gap shows up.
What is genuinely scarce in the open-source era is not yet another larger model, but someone willing to open-source the failing middle steps as well. Could this "reconstruct trajectories backward from finished videos" approach work for long documents or software delivery pipelines?