Plans You Can Check: Verifier-Grounded Learning of an Open-Weight Planner for Executable Video-Editing

Haoyu Wang; Cheng Feng; Liuyang Bian; Ruiyang Huang; Lei Wei; Yafei Wen; Xiaoxin Chen; Xiaoying Tang

cs.CVAug 26, 2026

Plans You Can Check: Verifier-Grounded Learning of an Open-Weight Planner for Executable Video-Editing

Authors: Haoyu Wang, Cheng Feng, Liuyang Bian, Ruiyang Huang, Lei Wei, Yafei Wen, Xiaoxin Chen, Xiaoying Tang

Organizations: School of Artificial Intelligence, The Chinese University of Hong Kong, Shenzhen · vivo AI Lab · University of the Chinese Academy of Sciences · Southeast University ·

Abstract

Practical video editing is not only pixel generation: an editor must turn a brief, a clip pool, music metadata, and hard constraints into an executable timeline. We study this decision layer as executable video-editing planning and introduce RefineCut, which, unlike workflow systems that wrap a prompted frontier model, trains a compact open-weight planner for it. The planner edits a typed timeline through structured patches covering clip selection, trimming, ordering, transitions, and duration and music alignment; a deterministic verifier applies each patch and checks it against an explicit constraint ledger. Because editing has no single ground-truth repair, we do not imitate teachers directly: RefineCut replays every multi-teacher branch through the verifier and keeps verifier-best repairs as supervision. A second stage, RefineCut-Evo, lets the student score its own repairs with the verifier and a task rubric and trains on high-margin preference pairs, so the final $8$ B planner runs in a closed verifier loop with no teacher calls at inference. On RefineCut-Bench ( $3{,}578$ tasks, $7{,}971$ captioned clips, $499$ music tracks, explicit ledgers), verifier-replayed distillation lifts the planner from $0.620$ to $0.858$ on the protocol-specific Video-Editing Score and RefineCut-Evo reaches $0.924$ ; the gain transfers to Llama-3.1-8B and GLM-4-9B, and in the same closed loop the $8$ B planner matches or exceeds its frontier teachers. Code and RefineCut-Bench are publicly released; see the Data Availability statement.

Figures & tables

Figure 1: From implicit to explicit planning. RefineCut turns implicit prompted editing into explicit, verifier-trained executable planning.

Figure 2: RefineCut-Bench overview. The benchmark ties real clip/music metadata, explicit ledger constraints, multi-teacher trajectories, and verifier-based evaluation into one planning-level protocol.

Figure 3: Framework. Noisy teacher trajectories become verified supervision; verifier- and rubric-scored student repairs drive self-improvement and closed-loop planning.

Figure 4: Training-stage progression. Common-100 VES across checkpoints; annotations mark the Raw

\to

Verified and Mixed-Pref

\to

Evo aggregate gains.

	Constraint adherence			Loop quality		Overall
Model	FCSR	Hard	ReqCR	Dur	Cvg	VES
Prompted backbone and raw imitation
Prompted	0.608	0.150	0.273	0.660	0.210	0.594
Raw	0.559	0.160	0.723	0.440	0.250	0.620
Verifier replay distillation
Verified	0.880	0.630	0.975	0.820	0.790	0.858

Table 1: Closed-loop planning results on Common-100. The

\Delta

row gives aggregate changes relative to Mixed-Pref; ⋆ marks the final RefineCut checkpoint.

	Replay scale		Verifier components
Teacher	$N$	Branches	FCSR	Hard	PASR	ReqCR	VES
GPT-5.4	2,000	6,000	0.379	0.015	0.872	0.698	0.521
Qwen3-Max	1,946	3,892	0.446	0.016	0.989	0.712	0.557
DeepSeek-V4-Pro	1,959	3,920	0.473	0.028	0.946	0.703	0.559

Table 2: Verifier-best replay diagnostics by teacher. Per-branch values are computed from replay logs; this diagnoses the teacher resource, not student closed-loop performance.

	Component metrics				Overall
Variant	FCSR	Hard	Dur	Cvg	VES
Rubric-margin DPO	0.9530	0.820	0.980	0.950	0.9237
Verifier-only DPO	0.9395	0.780	0.940	0.900	0.9085 $\downarrow$ 0.015
No-margin DPO	0.9178	0.760	0.910	0.870	0.8965 $\downarrow$ 0.027
UCPO-lite DPO	0.9012	0.740	0.890	0.840	0.8800 $\downarrow$ 0.044

Table 3: RefineCut-Evo ablation under matched backbone, candidate pool, DPO recipe, and wall-clock budget. Verifier-scored self-improvement is the primary signal; rubric-structured margins add

0.015

	VES by training stage			Gain
Backbone	Prompted	Raw	Verified	$\Delta$ V $-$ R
Qwen3-8B	0.594	0.620	0.858	$+0.238$
Llama-3.1-8B	0.566	0.496	0.649	$+0.153$
GLM-4-9B	0.643	0.607	0.686	$+0.079$

Table 4: Cross-backbone transfer. Same data, recipe, and frozen protocol; 95% CIs:

+0.153

[0.109,0.197]

+0.079

[0.036,0.120]

. Full components in Appendix G , Table 13 .

Policy (identical loop)	VES	Hard	Cvg
Frontier teachers (online)
GPT-5.4	0.893	0.76	0.86
Qwen3-Max	0.773	0.63	0.72
DeepSeek-V4-Pro	0.936	0.89	0.95
Untrained Qwen3-8B: prompting and search
Direct prompting ( $T{=}1$ )	0.502	0.07	0.08

Table 5: Same-loop external comparison. All policies use the identical state, ledger, RefinePatch interface, Apply/Verify loop, stopping rule, and

T{=}3

budget. Full tables with latency, W/T/L, CIs, and

R{=}1

sampling are in Appendix G .

Figure 5: Failure composition. RefineCut-Evo increases OK plans from 74 to 91 and reduces duration mismatches from 16 to 1 on Common-100; the small residual tail shows all remaining closed-loop failure types.

Appendix figures & tables5 assets

Supplementary material from the paper’s appendix.

Appendix

Item	Value
Raw video clips	7,971
Extracted caption frames	23,913
Music tracks	499
Raw task records	3,960
Canonical unique tasks	3,578
Train / Dev / Test (records)	2,773 / 596 / 591

Appendix

Table 6: Dataset, splits, and teacher coverage. Teacher counts are GPT-5.4 / Qwen3-Max / DeepSeek-V4-Pro.

Source	Captioned clips
panda70m_sample	3,990
pexels	3,219
pixabay	255
openvid	100
long_video_segment	407
Total	7,971

Appendix

Table 7: Captioned clip counts per source. The total matches the asset count reported in Section 2.4 .

Family	Subtypes
A (multi-clip	themed_montage, pacing_progression,
composition)	motion_dynamics_montage, scene_traversal,
	character_focus, abstract_rhythm
B (local	b_roll_insert, clip_swap, cut_extend,
edit/repair)	transition_repair
C (generative	generated_opener, generated_bridge,

Appendix

Table 8: The

15

task subtypes grouped by family.

repair_operator	Intent
modify_prompt	rewrite the generative prompt
change_tool	switch the downstream tool slot
adjust_parameter	tweak a numeric or enum field
reselect_clip	replace a clip in the sequence
retry_with_fallback	redo with a safer default
abort_and_skip	drop the step

Appendix

Table 9: The six repair_operator values.

ID	Criterion	Avg. weight
ER1	Intent and Story Alignment	0.09
ER2	Ledger Satisfaction	0.25
ER3	Clip Grounding and Relevance	0.22
ER4	Timeline and Segment Coherence	0.11
ER5	Duration and Pacing Alignment	0.15
ER6	Music / Beat Alignment	0.05

Appendix

Table 11: Average rubric weights across the train set (rounded; per-task weights sum to

1

). Family A puts more mass on ER4; families B and C put more mass on ER3.

Video Editing

Video editing uses AI to alter, compose, relight, or reorganize moving footage while controlling how changes persist across frames.

See all 179 papers

GRNEdit: Efficient General Video Editing from a New Binary-Evidence Perspective in Generative Refinement Networks
Aug 17, 2026Feng Xie, Jiagao Hu, Fuhao Li +5Video Editing Instruction-Based Video Editing
Xidian University · MiLM Plus, Xiaomi Inc.
Timeline-Bench: Evaluating Agents on Realistic Video-Editing Tasks, from Raw Footage to Final Cut
Sep 28, 2026Gunin Gupta, Nirmit Arora, Pavan Kalyan TankalaVideo Editing Benchmarks Agentic Video Editing
TensorTest (Ritivel Labs Inc.)
From Plans to Pixels: Learning to Plan and Orchestrate for Open-Ended Image Editing
May 14, 2026Anirudh Sundara Rajan, Krishna Kumar Singh, Yong Jae LeeText-Guided Image Editing Tool-Using Vision-Language Agents
University of Wisconsin–Madison, WI, USA · Adobe Research, San Jose, CA, USA

Aug 17, 2026J/K move · Enter open · S save

GRNEdit: Efficient General Video Editing from a New Binary-Evidence Perspective in Generative Refinement Networks

Feng Xie, Jiagao Hu, Fuhao Li, Zepeng Wang+4

Xidian University · MiLM Plus, Xiaomi Inc.

Video Editing Instruction-Based Video Editing

Open paper AskPDF arXiv

Instruction-based general video editing seeks to unify diverse editing operations within a single, intuitive interface. Existing approaches often rely on resource-intensive conditioning, using either heavyweight branches or costly source concatenation. Is there any efficient way to model editing intent? Thus, we introduce GRNEdit, a lightweight two-stage framework. GRN inspires our approach by encoding visual semantics through combinations of bits. Through task-specific fine-tuning, we take this representation further and recast editing semantics as local retain-or-flip decisions over individual bits. Source information is consequently modeled as coordinate-wise evidence supporting the observed binary states, while the GRN backbone remains responsible for resolving their global composition into coherent generative semantics. In Stage I, a compact encoder translates discrete source codes into continuous evidence signals, which GRN assimilates throughout binary refinement. Inspired by null-prompt training for classifier-free guidance, we further assign the null condition an editing-specific meaning: an empty instruction denotes no edit and is supervised through source reconstruction. This identity pathway not only implicitly strengthens evidence utilization and content preservation in Stage I, but also produces a source-preserving state in the same representation space as the edited state. Stage II can therefore directly compare each edited state with its source-preserving counterpart and use their discrepancy to revise unresolved target-bit decisions. Trained on only 0.6M pairs with less than 3% conditioning parameters, GRNEdit-2B and GRNEdit-8B achieve scores of 4.03 and 4.18 on OpenVE-Bench. The 2B model outperforms multiple 14B open-source editors, while the 8B model performs on par with leading open-source editors.

Abstract

Video Editing

Explore similar work

GRNEdit: Efficient General Video Editing from a New Binary-Evidence Perspective in Generative Refinement Networks

Timeline-Bench: Evaluating Agents on Realistic Video-Editing Tasks, from Raw Footage to Final Cut

From Plans to Pixels: Learning to Plan and Orchestrate for Open-Ended Image Editing