acceptodds
Under review as a conference paper at ICLR 2027

CutMaster: Towards Agentic Hour-Scale Video Editing

Abstract

Transforming an hour-scale source video into a short montage conditioned on background music, a user request, and a target duration requires balancing narrative alignment, emotional pacing, and visual quality. Existing agentic editing systems often organize decisions by processing stage or tool, leaving dependencies among these objectives implicit and resulting in suboptimal visual selection. We present **CutMaster**, an objective-factorized, stateful, and verification-driven agentic workflow built around a six-role team. A request-independent material analyst constructs reusable material memory, after which five request-conditioned roles coordinate their decisions through a shared editorial state. Music-aligned slots capture emotional pacing, story anchors provide narrative grounding, and a visually validated candidate space supports beam-search-based composition. Bounded replanning repairs intermediate states when earlier decisions become infeasible. We further introduce MontageX-Bench, a 40-task benchmark spanning four editing intents and three source domains. Compared with five recent state-of-the-art baselines, CutMaster achieves a Quality score of 81.40, outperforming the strongest baseline at 75.55, while providing the best overall balance across narrative, pacing, and visual-quality metrics.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
language agent, video editing

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_B1GvOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

CutMaster takes hour-long footage and edit it down into a short montage, based on a user request, a piece of background music, and target duration. Its workflow have six roles and bring together a reusable memory of the source, slots aligned to the music, dialogue anchors, candidate clips that is checked visually, and a sequence selection step with constraints. The paper also introduce MontageX-Bench, which has 40 task built from ten source video, and compares the system against five other editing systems.

Critical strengths

1.The workflow turn editing decisions into concrete steps you can actually inspect. Handling dialogue anchors, music-aligned slots, and candidate selection separately seem like a good way to deal with requirements that pulls against each other. The worked example also show visual verification catching a bad retrieval before anything get rendered. 2.The improvement is backed up by several separate check. CutMaster comes first in main evaluation, stay ahead when a second VLM judge is used, and get higher human ratings than the strongest baseline, 4.47 compared with 4.13. That's more convincing than a gain that only show up with single automated evaluator.

Critical weaknesses

1.Taking out the Timeline Scout and the Edit Composer together do show that these parts helps, but it doesn't isolate verification, sequence selection, or shared-state coordination on their own. The comparisons backs up an end-to-end advantage. To really support the central design claim, though, I'd want to see simplified workflow that use the same models and similar inference budget. 2.The variation in Table 1 come from three evaluation run, so it doesn't tell us how much the results varies across independently generated edits. The human study is useful, but the paper don't say how the 100 ratings are spread across systems and tasks, and it doesn't give the sample count or uncertainty behind pairwise preferences. Since everything comes from just ten source video, these details matters for judging how far the results generalize.

Questions for authors

1.If the models and inference budget are kept about the same, how much does verification and replanning actually adds compared with simplified workflow? It would help to also report how many model call were made, or what the cost was. 2.How many distinct task and ratings goes into each human score and each pairwise preference? Was the comparisons matched by task? And once you account for the same sources and participants showing up more than once, how much uncertainty are left in the differences?

4 comments
AuthorsOpenReview ↗

Response to Reviewer B1Gv (Part 1 of 2)

We are grateful for your constructive feedback and your positive recognition of our work. We address your concerns point by point below, and update the manuscript accordingly.

W1: Removing T and E together does not isolate their contributions; a simplified workflow comparison is needed.

Thanks for your detailed suggestion. Timeline Scout (T) retrieves candidate clips from the source video, defining the candidate space on which Edit Composer (E) operates; removing T alone would therefore leave E without its required input. We therefore introduce w/o E, which replaces E's global sequence selection with local clip selection.

We also introduce Simplified CutMaster workflow, which uses the same models in a single-pass workflow without VLM-based verification, scoring, retries, or replanning.

SystemInstruction FollowingNarrative CoherenceVisual QualityTransition ContinuityQuality
CutMaster4.63 ± 0.054.49 ± 0.034.77 ± 0.014.26 ± 0.0181.40 ± 0.10
w/o E4.55 ± 0.054.43 ± 0.044.71 ± 0.014.20 ± 0.0380.30 ± 0.39
w/o T&E4.49 ± 0.074.33 ± 0.034.68 ± 0.044.16 ± 0.0579.37 ± 0.06
Simplified4.44 ± 0.074.29 ± 0.084.46 ± 0.113.99 ± 0.0876.50 ± 0.42

Contribution of E. Removing E reduces Quality by 1.10 points, demonstrating the benefit of sequence-level optimization over local selection using the same candidates.

Contribution of T. Further removing T reduces Quality by 0.93 points, indicating the combined contribution of candidate retrieval and visual verification.

Simplified workflow. Compared with the full workflow (81.40), the simplified version achieves 76.50 Quality, with lower scores across all four VLM-judged metrics. This supports the benefit of verification and iterative repair over single-pass execution.

W2-1: Variation across independently generated edits.

To assess system stability, we rerun the benchmark twice with CutMaster, obtaining Quality scores of 80.69 and 80.67. Across all three generation runs, Quality is 80.96 ± 0.48, indicating consistent performance.

RunBCSAECIFNCVQTCQuality
183.1851.504.684.484.784.2581.51
283.5352.494.504.534.704.2080.69
383.6550.984.584.504.684.2380.67
Mean ± std83.45 ± 0.2551.66 ± 0.774.58 ± 0.094.50 ± 0.034.72 ± 0.054.23 ± 0.0380.96 ± 0.48

We have included these results in Appendix H.

W2-2 & Q2: The human study does not report how its 100 ratings are allocated across systems and tasks, or the sample counts and uncertainty behind pairwise preferences.

Thank you for the detailed suggestion. We clarify the human evaluation setup and statistical uncertainty below.

Human Likert scores. We randomly sampled 100 video outputs for Likert evaluation, covering 37 tasks and 10 source videos, with ratings from 10 participants. The resulting ratings comprise 17 for CutClaw, 16 for DIRECT, 16 for NarratoAI, 19 for OpenMontage, 17 for VideoAgent, and 15 for CutMaster.

Pairwise comparisons and uncertainty. All comparisons are task-matched. During rebuttal, we expanded the evaluation to 20 comparisons between CutMaster and each baseline. We report each baseline's preference rate as (wins + 0.5 × ties) / n, with approximate 95% Wilson confidence intervals. All baseline preference rates are below 50%, indicating a consistent preference for CutMaster.

Baseline (vs. CutMaster)Baseline winsTiesCutMaster winsBaseline Preference Rate95% Wilson CI
CutClaw11187.5%[1.8%, 26.9%]
DIRECT211712.5%[4.0%, 33.1%]
NarratoAI311617.5%[6.6%, 38.9%]
OpenMontage121710.0%[2.8%, 30.1%]
VideoAgent10195.0%[0.9%, 23.6%]

We have updated Figure 2 and included these analyses in Appendix I.

AuthorsOpenReview ↗

Response to Reviewer B1Gv (Part 2 of 2)

Q1: How much do verification and replanning improve over a simplified workflow?

We additionally evaluate a simplified CutMaster using the same models in a single-pass pipeline, without VLM-based verification or scoring, retries, or replanning. Here we report the API calls, API cost and runtime per task.

ConfigurationQualityAPI calls / taskAPI cost / taskRuntime / task
CutMaster81.40 ± 0.1063.78$0.525734.4 s
Simplified76.50 ± 0.4211.47$0.300223.0 s

The full workflow obtains a Quality score of 81.40, compared with 76.50 for the simplified pipeline, at an additional API cost of only $0.225 per task. Verification, scoring, and iterative repair increase inference expenditure while supporting higher-quality outputs. Moreover, the simplified workflow successfully produces final videos for only 27 of the 40 tasks, highlighting the importance of these mechanisms for reliable execution.

Dear Reviewer B1Gv,

As the discussion period draws to a close, we would appreciate your feedback on whether our rebuttal has addressed your concerns. We would be happy to clarify any remaining issues.

Best regards,

The Authors

Thanks you for the detailed response. I decided to keep my positive score.

Reviewer_AkqFOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper presents CutMaster, a multi-agent workflow for converting hour-scale source videos into short montages. The workflow consists of six roles (Material Analyst, Arrangement Architect, Story Editor, Timeline Scout, Edit Composer, and Renderer) that communicate through a versioned Shared Editorial State. The paper also introduces MontageX-Bench, consisting of 40 tasks across 10 long videos, four editing intents, and three source domains.

Critical strengths
  1. Transforming an hour-scale source video into a short montage is a meaningful and challenging video editing task.
  2. The design of the proposed workflow is clear. Six-role organization is easy to understand, and the responsibilities of the roles are reasonably well separated.
Critical weaknesses
  1. The paper motivates the six-role workflow intuitively, but does not provide a principled justification for this particular decomposition. It remains unclear why six roles are necessary or why these role boundaries are preferable to alternative groupings.
  2. The baseline comparisons do not sufficiently control computational resources and editing policies. Appendix E explicitly states that the comparison does not enforce equal model backbones, compute budgets, or rendering policies. It is therefore difficult to determine how much of the gain comes from the workflow design versus stronger models, more model calls, or more extensive engineering.
  3. The scalability, efficiency, and generalization capabilities of the framework remain insufficiently explored. CutMaster is designed for hour-scale video editing, but the experiments do not report end-to-end processing time, memory requirements, API costs, or the number of replanning iterations. Additionally, the benchmark contains only 10 source videos and 11 background music tracks.
Questions for authors
  1. Why is beam search necessary for the proposed sequence optimization problem? Given that Equation 6 contains only unary candidate scores and adjacent pairwise scores, could the authors explain why the problem cannot be solved exactly using dynamic programming?
  2. The four Likert metrics are scored by Qwen3.7-Plus, and CutMaster uses the same model internally for shot annotation, candidate verification (subject visibility and visual relevance), and pairwise transition scoring. This means the system is in effect optimized against the same model that later judges it, which may favor CutMaster’s outputs. Could the authors discuss this potential self-preference bias?
  3. The paper says CutMaster differs from prior systems by organizing decisions around objectives rather than processing stages. However, each MASTER role still corresponds to a sequential stage (planning, anchoring, retrieval, composition, rendering). Music-anchored planning, retrieval, and review also already appear in systems such as CutClaw and DIRECT. Could the authors state more precisely which mechanisms in CutMaster are not present in these systems and proof which of them account for the performance gains.
  4. The problem formulation requires the montage to follow the original source chronology. This is reasonable for event-oriented requests. For emotion- or character-oriented montages, however, professional editors often reorder material to build a stronger emotional arc. Could the authors explain the motivation for this constraint?
4 comments
AuthorsOpenReview ↗

Response to Reviewer AkqF (Part 1 of 2)

We sincerely appreciate your constructive feedback. Below we respond to your concerns point by point, and we have accordingly revised the manuscript.

W1: It remains unclear why six roles are necessary or why these role boundaries are preferable to alternative groupings.

The six roles represent CutMaster's decomposition of the long-video editing task.

  • Material analysis and final rendering are required before and after editing, respectively, so we assign the Material Analyst (M) and Renderer (R) to handle preparation and execution.
  • The intermediate editing-planning stage is handled by the Arrangement Architect (A), Story Editor (S), Timeline Scout (T), and Edit Composer (E). The decomposition principle is to separate the three optimization objectives—emotional pacing, narrative, and visual quality. A handles emotional pacing, S handles narrative, and T and E handle visual quality. The ablation results in Table 1 provide empirical support for this division of responsibilities.

W2: The baseline comparisons do not sufficiently control computational resources and editing policies.

We clarify that our experiments use the same models for corresponding functions across systems: Qwen3.7-Max for LLM reasoning, Qwen3.7-Plus for visual understanding, and Bailian's Fun-ASR for transcription, as reported in Table 9. We impose no additional compute limits, allowing each system to follow its own workflow. We have revised Appendix E to avoid misunderstanding.

Given the same selected models, CutMaster's higher score in Table 1 supports the effectiveness of the proposed workflow.

W3-1: Processing time, memory requirements, API costs and the number of replanning iterations.

We report editing-stage API costs and cumulative runtimes over the 40 tasks below. These results highlight a practical quality–efficiency trade-off. CutMaster is not the lowest-cost system, but it is faster and less expensive than CutClaw and DIRECT-Claw, while achieving stronger quality results.

SystemAPI cost (USD)Cumulative runtime (h)GPU memory requirement
CutClaw$66.534.63Not required
DIRECT$23.419.5424 GB GPU used for retrieval component
OpenMontage$16.110.06Not required
CutMaster$21.08.16Not required

Across the 40 tasks, CutMaster performed 29 local replanning iterations and 4 global replanning iterations. In local replanning, the Arrangement Architect revises individual slot descriptions, whereas global replanning regenerates the slot plan based on accumulated failure feedback.

W3-2: The benchmark contains only 10 source videos and 11 background music tracks.

Since the benchmarks used by related video editing systems are not publicly available, we constructed MontageX-Bench from scratch: each of ten source videos is paired with four editing intents, producing 40 tasks across sports, documentaries, and films. As shown in Table 6, MontageX-Bench is comparable to prior work in task count and covers multiple content domains.

BenchmarkTasksSource videosDomains
CutClaw2010Film, daily life
DIRECT4038Film
Crayotter23— (Not reported)Pet, campus, travel, scenery, food
MontageX-Bench4010Sports, documentary, film
AuthorsOpenReview ↗

Response to Reviewer AkqF (Part 2 of 2)

Q1: Why is beam search necessary for the proposed sequence optimization problem?

We agree that the exact dynamic programming can solve Eq. 6 when the scores are available and the feasibility constraints are encoded in the state or transitions. However, we use beam search to reduce the cost of obtaining pairwise scores. With a beam width of 8 and candidates number of 3, we request VLM scores only for transitions explored from retained paths. This trades exact optimality for 44.96% fewer scoring requests across 40 tasks.

Pairwise VLM scoring across 40 tasksRequests
Exhaustive scoring of adjacent candidate pairs1,835
Beam search (width 8, candidate 3)1,010
Reduction44.96%

Q2: Discuss the potential self-preference bias.

Thank you for raising this concern. To examine potential evaluator bias, Appendix B evaluates the same generated videos with GLM-5.3-Flash, using the same Likert rubric. In Table 7, CutMaster remains the highest-scoring system in Instruction Following, Narrative Coherence, and Transition Continuity, and ties for the highest Visual Quality. It also retains the highest overall Quality, consistent with the Qwen3.7-Plus evaluation. These results suggest that the observed advantage is not specific to Qwen3.7-Plus.

SystemBeat–Cut Sync.Energy Corr.Instruction FollowingNarrative CoherenceVisual QualityTransition ContinuityQuality
CutClaw57.3356.873.253.534.303.8564.57
DIRECT73.9857.833.203.183.683.5361.97
NarratoAI57.2452.924.034.304.784.4374.78
OpenMontage56.1249.503.513.644.153.8564.06
VideoAgent65.6154.393.253.384.483.8365.58
CutMaster83.1851.504.554.584.784.5382.61
w/o T&E83.0451.724.404.534.784.4081.25
w/o S86.7154.994.234.354.634.4080.33

Q3: Could the authors state more precisely which mechanisms in CutMaster are not present in these systems?

Thank you for the constructive feedback. CutMaster's key differences lie in how editing decisions are constrained, optimized, and refined across stages:

  • Narrative anchoring. Unlike conventional music-driven planning, CutMaster anchors key dialogue to corresponding footage, preserving narrative constraints during editing. Removing the Story Editor reduces Narrative Coherence by 0.22 and overall Quality by 2.51.
  • Sequence-level selection. Unlike CutClaw's clip-level selection and DIRECT's segment-level retrieval, CutMaster jointly optimizes clip selection across the entire montage using verified candidates. Replacing this with local selection decreases overall Quality by 1.10.
  • State-driven repair. Unlike CutClaw's local retrieval feedback or DIRECT's query adjustment, CutMaster uses shared slots, anchors, candidates, and diagnostics to revise affected upstream decisions. The simplified pipeline without verification and iterative repair reduces overall Quality from 81.40 to 76.50.

We have clarified these distinctions in Related Work and their empirical support in the Ablation Study.

Q4: The method follows the original source chronology, while professional editors often reorder material to build a stronger emotional arc.

Thank you for raising this point. We agree that reordering footage can enhance character- or emotion-oriented montages. In our single-source long-video editing setting, we preserve source chronology to maintain the original context of events and dialogue and avoid misleading causal interpretations. We acknowledge that this constraint may limit editing flexibility and have added it in Limitations.

Dear Reviewer AkqF,

As the discussion phase is coming to a close, we would appreciate your feedback on whether our rebuttal has addressed your concerns. We would be happy to clarify any remaining issues.

Best regards,

The Authors

Thank you for the detailed response and the additional analyses. The clarifications and new results are helpful for understanding the system, and I appreciate the effort the authors have put into the rebuttal. After considering the response and the other reviews, I have decided to keep my score.

Reviewer_bjH7OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
3(+1)
Read review
Paper summary

The paper introduces CutMaster, an objective-factorized agentic workflow designed for hour-scale source-to-short video montage editing. Addressing the challenge of balancing narrative alignment, emotional pacing, and visual quality, CutMaster utilizes a six-role "MASTER" team (Material Analyst, Arrangement Architect, Story Editor, Timeline Scout, Edit Composer, and Renderer) coordinated through a versioned Shared Editorial State. Additionally, the authors also introduce MontageX-Bench, a 40-task evaluation suite that spans four editing intents (event, character, emotion, narrative) and three source domains (sports, documentaries, films). Against five recent baselines, CutMaster achieves a state-of-the-art Quality score of 81.40, outperforming the strongest baseline (NarratoAI) at 75.55.

Critical strengths
  • The task is useful and challenging: editing hour-long footage into a short video while following a request and matching music.
  • The benchmark tests four editing intents on each of ten source videos, which makes the comparisons across intents more informative.
  • The system achieves strong results across the benchmark, and the human evaluations generally support the automated results.
Critical weaknesses
  • The writing is not very good, especially the related work section. In lines 97–107, I honestly cannot tell what the authors are trying to say. They list papers on different tasks and then jump to a broad claim about what those papers assume, without explaining the connection. It reads like an LLM-generated list of citations but not a meaningful discussion of prior work.
  • The paper is incremental compared to CutClaw. CutClaw already formulates essentially the same problem. CutClaw also analyzes the footage, plans around the music, retrieves clips, and verifies them. CutMaster adds dialogue anchors and sequence-level selection, but I do not think these changes make its first claimed contribution to the problem formulation and framework sound and substantial.
  • I am not entirely convinced by the current experiments. The systems use different models, compute budgets, and rendering policies. CutMaster’s higher score over CutClaw therefore cannot entirely be attributed to its proposed workflow.
  • The ablations miss some important ones. For example, “w/o T&E” removes the Timeline Scout and Edit Composer together, so it cannot tell whether the gain comes from candidate verification, having multiple candidates, or sequence-level selection. “w/o S” only tests dialogue anchors. I will also be interested in learning more about controlled comparisons between the proposed factorized workflow and a single editor using the same tools and budget, or global sequence selection versus local clip selection using the same candidates.
Questions for authors
  1. What is the main methodological difference from CutClaw, beyond adding more roles? Can you test it with the same models and compute budget?
  2. In “w/o T&E,” which matters: visual verification, multiple candidates, or sequence-level selection? Can you ablate them separately?
  3. Given the unary and adjacent-pair scores in Eq. 6, why use beam search instead of exact dynamic programming?
  4. How many comparisons underlie each human preference percentage in Figure 2, and what are the confidence intervals?
6 comments
AuthorsOpenReview ↗

Response to Reviewer bjH7 (Part 1 of 2)

We sincerely appreciate your constructive feedback. Below we address your concerns point by point, and we have accordingly updated the manuscript.

W1: The writing is not very good, especially the related work section.

Thank you for the constructive feedback. In the revised manuscript, we clarify what each line of research contributes and how our approach builds on or differs from it.

W2 & Q1-1: The paper is incremental compared to CutClaw. What is the main methodological difference from CutClaw, beyond adding more roles?

Thank you for the constructive feedback. CutMaster shares CutClaw’s music-driven, hour-scale editing setting but differs in three aspects:

  • Different motivation: CutClaw focuses on long-video understanding and audiovisual alignment, while CutMaster focuses on organizing and revising decisions when narrative, emotional pacing, and visual quality conflict.
  • Different framework: CutClaw verifies and commits clips individually under a music-driven script, whereas CutMaster retains multiple validated candidates for joint sequence selection, guided by music-aligned slots and dialogue anchors.
  • Different implementation: CutClaw uses feedback to search for alternative clips locally, whereas CutMaster uses shared editorial state to revise upstream slots and coordinate affected downstream decisions when the plan becomes infeasible.

These differences concern how editorial decisions are coordinated, selected, and revised, beyond the number of agent roles. We clarify these distinctions in the updated Related Work section.

W3 & Q1-2: The systems use different models, compute budgets, and rendering policies. Can you test it with the same models and compute budget?

We clarify that our experiments use the same models for corresponding functions across systems: Qwen3.7-Max for LLM reasoning, Qwen3.7-Plus for visual understanding, and Bailian's Fun-ASR for transcription, as reported in Table 9. We impose no additional compute limits, allowing each system to follow its own workflow. We have revised Appendix E to avoid misunderstanding.

Under the same language and vision backbones, CutMaster's higher score in Table 1 supports the effectiveness of the proposed workflow. We report editing-stage API costs and cumulative runtimes over the 40 tasks below. CutMaster requires less time and API expenditure than CutClaw.

W4-1 & Q2: Can you ablate "w/o T&E" separately? Controlled comparisons between global sequence selection versus local clip selection using the same candidates.

Thanks for your detailed suggestion. Timeline Scout (T) retrieves candidate clips from the source video, defining the candidate space on which Edit Composer (E) operates; removing T alone would therefore leave E without its required input. We therefore only add w/o E, which replaces E's global sequence selection with local clip selection.

SystemBCSAECIFNCVQTCQuality
CutMaster83.1851.504.63 ± 0.054.49 ± 0.034.77 ± 0.014.26 ± 0.0181.40 ± 0.10
w/o T&E83.0451.724.49 ± 0.074.33 ± 0.034.68 ± 0.044.16 ± 0.0579.37 ± 0.06
w/o E82.9351.784.55 ± 0.054.43 ± 0.044.71 ± 0.014.20 ± 0.0380.30 ± 0.39

Contribution of E. Removing E reduces Quality from 81.40 to 80.30, a drop of 1.10 points. With the same candidate pool, this supports the benefit of sequence-level selection based on both clip quality and adjacent transitions.

Contribution of T. Further removing T lowers Quality to 79.37, another drop of 0.93 points. This supports T's role in providing better clip choices through candidate retrieval and visual verification.

W4-2: Controlled comparisons between the proposed factorized workflow and a single editor using the same tools and budget.

OpenMontage provides a complementary single-agent baseline: a coding agent (i.e. Claude Code) edits through skills and tools, using Qwen3.7-Plus as its underlying reasoning model. In Table 1, CutMaster achieves 81.40 Quality, compared with 62.19 for OpenMontage, supporting the effectiveness of structured editorial coordination over a general-purpose single-agent.

AuthorsOpenReview ↗

Response to Reviewer bjH7 (Part 2 of 2)

Q3: Why use beam search instead of exact dynamic programming?

We agree that the exact dynamic programming can solve Eq. 6 when the scores are available and the feasibility constraints are encoded in the state or transitions. However, we use beam search to reduce the cost of obtaining pairwise scores. With a beam width of 8 and candidates number of 3, we request VLM scores only for transitions explored from retained paths. This trades exact optimality for 44.96% fewer scoring requests across 40 tasks.

Pairwise VLM scoring across 40 tasksRequests
Exhaustive scoring of adjacent candidate pairs1,835
Beam search (width 8, candidate 3)1,010
Reduction44.96%

Q4: How many comparisons underlie each human preference percentage in Figure 2, and what are the confidence intervals?

Thanks for your insightful suggestion. We have updated Figure 2 in the revised manuscript to include the number of pairwise comparisons.

During rebuttal, we expanded the evaluation to 20 comparisons between CutMaster and each baseline. We report each baseline's preference rate as (wins + 0.5 × ties) / n, with approximate 95% Wilson intervals. All baseline preference rates are below 50%, indicating a consistent preference for CutMaster.

Baseline (vs. CutMaster)Baseline winsTiesCutMaster winsBaseline Preference Rate95% Wilson CI
CutClaw11187.5%[1.8%, 26.9%]
DIRECT211712.5%[4.0%, 33.1%]
NarratoAI311617.5%[6.6%, 38.9%]
OpenMontage121710.0%[2.8%, 30.1%]
VideoAgent10195.0%[0.9%, 23.6%]

Dear Reviewer bjH7,

As the discussion phase is coming to a close, we would appreciate your feedback on whether our rebuttal has addressed your concerns. We would be happy to clarify any remaining issues.

Best regards,

The Authors

Comments on the paper

2 comments
AuthorsOpenReview ↗

General Response and Summary of Revisions

Dear Reviewers and Area Chair,

We sincerely thank you for your time and constructive feedback. We appreciate the reviewers’ recognition that:

  1. The task is meaningful and challenging [bjH7, AkqF], and the presentation is clear and easy to follow. [AkqF, B1Gv]
  2. The workflow design is clear, and the responsibilities of the roles are reasonably well separated. [AkqF]
  3. The benchmark makes the comparisons more informative, and the system achieves strong results across the benchmark. [bjH7]

Following your suggestions, we have made the following revisions in the updated manuscript (highlighted in red):

  1. Revised Related Work to clarify the method differences between CutMaster and prior systems. Section 2 [bjH7, AkqF]
  2. Added w/o E and Simplified ablations to evaluate the effectiveness of sequence-level selection and the contribution of verification and iterative repair. Section 4.3, Table 1 [bjH7, AkqF, B1Gv]
  3. Added two independent generation runs, showing consistent gains over baselines. Appendix H [B1Gv]
  4. Expanded human preference evaluation, with 20 pairwise comparisons for each baseline–CutMaster pair. Section 4.4, Figure 2, Appendix I. [bjH7, B1Gv]
  5. Clarified baseline settings, including the same LLM/VLM underlying model in all experiments. Appendix E [bjH7, AkqF]
  6. Added limitations concerning source chronology. Section 5 [AkqF]

We sincerely appreciate your constructive suggestions, which help strengthen our paper a lot. We are happy to provide additional clarifications. Thank you for your time!

Sincerely,
The Authors

AuthorsOpenReview ↗

Summary of Main Contributions and Rebuttal Outcomes

Dear Area Chair and Reviewers,

We sincerely appreciate your time and efforts in handling our submission. To facilitate your final assessment, we summarize the main contributions of our work and the key outcomes of rebuttal as follows:


Part I: Main Contributions

  1. Multi-objective editing. For music-driven editing requests, CutMaster introduces an objective-factorized, stateful, and verification-driven workflow. It makes dedicated decisions and revises for emotional pacing, narrative alignment, and visual quality.
  2. Comprehensive benchmark with reliable evaluation. MontageX-Bench pairs ten source videos with four editing intents yielding 40 tasks, and evaluates the results using six metrics covering narrative alignment, music-conditioned pacing, and visual quality.
  3. Consistent gains across intents and domains. CutMaster scores 81.40 in Quality, 5.85 points above the strongest of five baselines, and leads across all intents and domains. Human evaluators also prefer CutMaster over every baseline.

Part II: Key Outcomes of Rebuttal

  1. Reviewer bjH7 (Rating 2, Soundness 2 → 3)

    • [Review] Asked how CutMaster differs from CutClaw and requested a controlled ablation.
    • [Response] We revised the manuscript to highlight CutMaster's contributions and differences from CutClaw, added the w/o E ablations, and clarified the shared model setting.
    • [Feedback] The reviewer said our rebuttal "addressed my main soundness concerns," raised soundness to 3, and indicated they would raise the overall rating to 3 after revision.
    • [Follow-up] We revised the Introduction to clarify the motivation and highlighted the main takeaways in the Experiments.
  2. Reviewer AkqF (Rating 2)

    • [Review] Asked about role specialization, model settings, efficiency, and the stability of VLM-based evaluation.
    • [Response] We explained how roles divide the editing objectives, clarified the shared model setting, reported runtime and API costs and showed that CutMaster also ranks first with a second VLM judge.
    • [Feedback] The reviewer found the clarifications and new results "helpful for understanding the system" and maintained their rating.
  3. Reviewer B1Gv (Rating 3)

    • [Review] Requested a simplified workflow comparison, independent runs, and details on human preferences.
    • [Response] We reported a 4.90-point gain over the simplified workflow, consistent results across three runs, and confidence intervals for all five human preference comparisons.
    • [Feedback] The reviewer "decided to keep the positive score."

Best,
The Authors

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.