A compact 8B open-weight planner trained through RefineCut-Evo reached a VES of 0.924 in the study's closed-loop test, compared with 0.620 for raw imitation. In the same Apply/Verify loop, it scored above the GPT and Qwen teacher branches and statistically tied the DeepSeek teacher branch.
A planner built around checks
RefineCut treats editing as executable planning over a typed timeline. The planner emits structured RefinePatch JSON patches; a deterministic verifier applies them and recomputes an explicit constraint ledger, a running record of the brief's requirements. The runtime can repeat the Apply/Verify step for at most three repairs.
The training was offline and had two stages. First, verifier-replayed teacher trajectories were used to train an initial 8B planner. Then RefineCut-Evo sampled four repairs for each state, scored them with seven rubric criteria, ER1 through ER7, and trained on high-margin preference pairs. At inference, the Apply/Verify loop allowed at most three repair steps.
Why the first traces were weak
A diagnostic result helps explain the training design. The verifier-best replay branches had patch-apply success rates, or PASR, from 0.87 to 0.99, yet their HardPass rate was no higher than 0.03. In other words, branches usually produced something the system could apply, but rarely met the full ledger in one step; the paper therefore treats first-choice teacher traces as weak labels.
A better score after repair training
The reported VES values were 0.620 for raw imitation, 0.858 after verifier-replayed SFT, 0.864 with Mixed-Pref and 0.924 with RefineCut-Evo. The component measures moved in the same direction: HardPass rose from 0.670 to 0.820, DurationPass at the 2-second horizon from 0.830 to 0.980, and Converged@3 from 0.800 to 0.950.
On the 100 Common-100 test tasks, Evo scored higher than Mixed-Pref on 22, lower on five and tied on 73. Its mean paired VES difference was +0.059, with a 95% confidence interval from +0.028 to +0.092.
The result traveled across models
The transfer result appeared across backbones. Verified SFT exceeded raw imitation by +0.238 VES on Qwen, +0.153 on Llama and +0.079 on GLM. The reported 95% confidence intervals were [+0.109, +0.197] for Llama and [+0.036, +0.120] for GLM.
In the matched frontier comparison, Evo's paired VES difference was +0.030 against the GPT teacher, with a 95% interval of [+0.001, +0.062], and +0.150 against Qwen, with [+0.091, +0.213]. Against DeepSeek it was -0.012, with an interval of [-0.034, +0.010], which the study classified as a statistical tie.
Human checks, with the result still tied to the benchmark
The authors also checked whether that ordering held in more human-facing tests. On Human50, which contained 50 free-form briefs, Evo's VES was 0.902 versus 0.848 for Mixed-Pref. In blind randomized rendered-preview comparisons, 150 pairs were judged by three annotators per pair: Evo versus Mixed produced 100 wins, 34 ties and 16 losses for Evo, a preference score of 0.780 and agreement kappa of 0.620. Mixed-Pref versus Prompted had a preference score of 0.887.
The study also reports that changing input semantics was associated with lower scores. With all clip semantics removed, VES was 0.8502 versus 0.9237 in the unmodified control. Shuffling semantic descriptions was associated with VES of 0.7241, while required-clip recall fell from 0.98 to 0.46.
A defined test and a public release
RefineCut-Bench contains 3,578 canonical tasks built from 7,971 captioned clips, 23,913 caption-anchored frames and 499 music tracks, across three task families and 15 subtypes. Main testing used Common-100 with a shared PatchPlanner prompt; a canonical-clean subset contained 92 records, and the step-300 checkpoint was selected on dev100 before test measurement. No teacher was called at inference.
These figures are best read as performance under the stated typed-timeline and ledger contract and test protocol.
The paper states that its code and RefineCut-Bench are publicly available, with the release described as including benchmark tasks, metadata, canonicalized trajectories, replay scores and schemas. The manuscript is an arXiv v1 preprint dated 26 Aug 2026.
Paper data and sources
Original title: Plans You Can Check: Verifier-Grounded Learning of an Open-Weight Planner for Executable Video-Editing
Authors: Haoyu Wang, Cheng Feng, Liuyang Bian et al.
Journal/Repository: arXiv
Status: Preprint, not yet peer-reviewed
First online: 2026-08-26
DOI: Not available
Original paper · Full text