acceptodds
Under review as a conference paper at ICLR 2027

ReVi: Recursive Recap Video Generation

Abstract

Novels tell rich stories, but their textual form cannot directly present the sights and sounds of the worlds they describe. Film adaptations offer an audiovisual experience but require substantial production effort and viewing time; film recaps are shorter but depend on existing footage. Neither route directly provides a concise audiovisual account of a novel. We therefore study novel-to-recap video generation: producing a self-contained, source-grounded audiovisual retelling directly from a novel within a fixed viewing-time budget. This task requires compressing the story while preserving causal connections, allocating sufficient time to retained events, and aligning narration with newly generated visuals. We propose ReVi, a dual-loop multi-agent framework for Recursive Recap Video Generation. Its Inner Loop plans story content, narration time, and corresponding shots before video generation. Across novels, its Outer Loop performs recursive self-improvement by revising both the editing policy and the meta policy that governs its revision. Experiments on two evaluation sets show that ReVi improves recap quality and duration compliance over representative baselines, with human ratings and multi-round improvement results providing further support. Our code is available at https://anonymous.4open.science/r/ReVi-84EF.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
LLM Agents, Recap Generation

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_pM8WOpenReview ↗
Rating
3(+1)
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper studies novel-to-recap video generation, aiming to produce self-contained recap videos with narration and generated visual content directly from novels under a prescribed duration constraint. To address the challenges of long-form narrative compression, duration allocation, and audiovisual alignment, the authors propose ReVi, a multi-agent framework consisting of an inner loop for video production and an outer loop for recursive policy improvement. In the inner loop, Story Mapper constructs a structured StoryGraph, Paper Editor selects and reorganizes narrative events, Radio Editor generates narration and uses text-to-speech (TTS) feedback to estimate actual speech duration, and AV Planner converts narration into dramatic beats and corresponding video-generation requests. The framework also incorporates error diagnosis and stage-specific repair. In the outer loop, ReVi iteratively revises both the editing policy and the meta-policy responsible for diagnosing failures, proposing policy updates, and allocating search resources. Experiments on FilmEval and Recap32 demonstrate improvements in recap quality and duration compliance over several baseline methods.

Critical strengths

The paper addresses a well-defined generation problem with practical relevance. Generating a concise audiovisual recap directly from a long novel requires coordinating narrative understanding, content selection, narration duration, and visual planning. The explicit duration constraint makes the task more concrete than unconstrained text-to-video generation.

Rather than directly prompting a video generator with the entire novel, ReVi separates narrative planning, editing, narration, and visual generation into distinct stages. The use of measured TTS duration, dramatic-beat-based planning, and stage-specific repair provides a coherent workflow for addressing different types of generation failures.

The experimental evaluation covers several relevant aspects of the framework. In particular, the component ablations provide preliminary evidence that duration feedback, beat planning, and repair routing contribute to system performance. The outer-loop experiments also suggest that revising reusable editing policies can improve generation quality beyond using a frozen policy.

Critical weaknesses

The novelty of the proposed components is not sufficiently clear. StoryGraph, structured plot editing, TTS-based duration adjustment, and dramatic-beat planning are reasonable design choices, but the paper does not clearly distinguish them from existing approaches to summarization and agent-based video generation. The authors should clarify what new technical capabilities these components provide beyond integrating existing techniques.

The outer-loop optimization is the most interesting part of the paper, but its contribution remains unclear. Promptbreeder and GEPA have already explored iterative prompt optimization and feedback-driven improvement. The authors should explain what distinguishes ReVi's dual-loop mechanism from these methods and what specific advantage the meta-policy provides.

The outer-loop experiments do not adequately establish the benefit of the proposed optimization strategy. The two optimization settings evaluate different numbers of candidates (147 versus 195), making it difficult to separate the contribution of the meta-policy from the effect of a larger search budget. A comparison with an existing prompt optimizer under comparable evaluation budgets would provide stronger evidence.

The use of the Recap32 test novels in the outer-loop experiments needs clarification. Although N01–N12 are designated as test novels, Appendix D.3 reports iterative search results on six of them. The authors should clarify whether these results influenced policy selection or promotion. If they did, the final evaluation may not be independent.

The baseline comparison does not isolate the contribution of recursive policy optimization. ReVi achieves 100% duration compliance, while all reported baselines achieve 0%, but ReVi also employs explicit duration allocation, TTS feedback, and repair. A comparison under a shared duration-control setup would help determine how much improvement comes from these mechanisms versus the proposed optimization framework. The reported cost analysis should also account for offline policy-search costs.

Questions for authors

1、What are the specific methodological differences between ReVi's dual-loop optimization and existing approaches such as Promptbreeder and GEPA? Which components of the optimization procedure constitute genuinely new mechanisms rather than task-specific adaptations? 2、Does the improvement from meta-policy optimization remain when ReVi is compared with an existing prompt optimizer under the same candidate-evaluation budget and comparable computational costs? 3、Were N02, N04, N06, N08, N10, and N12 used exclusively for evaluation, or did their results influence policy selection or promotion during the outer-loop experiments? How was the independence of the final test evaluation maintained? 4、How were the automatic evaluation metrics validated? Can the authors provide additional human-evaluation results to establish the reliability of the reported quality improvements? 5、After including the computational costs of offline policy search and meta-policy evaluation, does ReVi retain a meaningful quality–cost advantage over the baselines?

11 comments
AuthorsOpenReview ↗

Response to Reviewer pM8W [1/7]

We sincerely thank the reviewer for the careful reading and the constructive comments. We thank the reviewer for recognizing the practical relevance of the novel-to-recap task, the coherent separation of narrative planning, editing, narration, and visual generation, the use of measured TTS duration, dramatic-beat planning, and stage-specific repair, and the evidence from both the component ablations and the outer-loop experiments.

[W1] The novelty of the proposed components is not sufficiently clear. StoryGraph, structured plot editing, TTS-based duration adjustment, and dramatic-beat planning are reasonable design choices, but the paper does not clearly distinguish them from existing approaches to summarization and agent-based video generation. The authors should clarify what new technical capabilities these components provide beyond integrating existing techniques.

Response: We thank the reviewer for this comment. We agree that the paper should state what each component makes possible, not only how it is built. Existing summarization methods decide what to keep, and existing agentic video pipelines decide how to show a given script. Novel-to-recap generation requires deciding what to keep, how long to tell it, and what to show jointly under a fixed runtime, and each ReVi component supplies one capability that this joint decision needs. We address the concern from two perspectives.

(a) Capability comparison with existing systems. The table compares ReVi with all baselines of the main comparison. For ReVi, the mechanism that provides each capability is given in parentheses.

MethodMulti-stage agentic planningSource-linked event structureCausal event selectionTime controlEvent-aligned shot planningRoot-cause repairCross-story self-improvement
Direct LLM✗✗✗✗✗✗✗
MM-StoryAgent✓✗✗✗✗✗✗
VGoT✓✗✗✗✗✗✗
VideoClaw✓✗✗✗✗✗✗
MovieAgent✓✗✗✗✗✗✗
ViMax✓✗✗✗✗✗✗
ReVi✓ (Story Mapper → three editors)✓ (StoryGraph)✓ (Paper Editor, causal closure)✓ (Radio Editor, measured TTS and narration modes)✓ (AV Planner, Dramatic Beats)✓ (cross-stage repair routing)✓ (dual-loop editing and meta policies)

The capabilities are defined as follows.

  • Source-linked event structure. Events with their causal, prerequisite, and setup–payoff edges, recorded in both chronological and disclosure order [Bordwell, 2008], with every downstream artifact linked back to its source passages.
  • Causal event selection. Under a fixed budget, no retained event loses its causes or setups; the Paper Editor applies six operations (Keep, Compress, Omit, Merge, Bridge, Externalize) and a fixed checker [Trabasso & van den Broek, 1985].
  • Time control. The produced runtime is measured from the synthesized speech and fitted to the target before rendering, with time allocated by narrative function (ellipsis, summary, scene, pause) [Genette, 1980]. All baselines have a 0% duration-pass rate on both datasets (Table 1).
  • Event-aligned shot planning. Shots follow perceptible event boundaries [Zacks et al., 2007] rather than sentences or scenes, with narration carrying internal states and visuals carrying observable change.
  • Root-cause repair. A failure observed in the video is routed to the earliest responsible stage (source, selection, timing, or visual planning) rather than revised where it is reported.

Hierarchical summarization methods (e.g., BookSum-style chapter-to-book summaries) share the first step with ReVi, but they select by salience, with no runtime budget and no check that the retained events remain causally closed.

(b) Each capability has its own measurable effect. The component ablations already reported in Appendix D.2 remove one capability at a time on the same six sources:

Removed capabilityAblationLargest loss
Causal event selectionw/o Causal EditingRF −4.77
Time allocation by narrative functionw/o Narrative AllocationNP −5.35
Time control from measured speechw/o TTS FeedbackDP 6/6 → 0/6
Event-aligned shot planningw/o Beat PlanningAVG −8.93
Root-cause repairw/o Repair RoutingRF −4.64, NP −3.10, AVG −3.17

As shown in the table, each removal costs the dimension that the component is designed to protect: causal selection protects recap fidelity, time allocation protects pacing, beat planning protects audio-visual grounding, and measured speech protects the duration contract. These capabilities are therefore not interchangeable prompt stages, and their combination is what enables the joint decision under a fixed runtime.

AuthorsOpenReview ↗

Response to Reviewer pM8W [2/7]

[W2] The outer-loop optimization is the most interesting part of the paper, but its contribution remains unclear. Promptbreeder and GEPA have already explored iterative prompt optimization and feedback-driven improvement. The authors should explain what distinguishes ReVi's dual-loop mechanism from these methods and what specific advantage the meta-policy provides.

Response: We thank the reviewer for finding the outer loop the most interesting part of the paper. The concerns center on how ReVi's dual loop differs from Promptbreeder and GEPA, and what the meta policy adds. We try to address each concern below.

W2-1. Promptbreeder and GEPA have already explored iterative prompt optimization and feedback-driven improvement. The authors should explain what distinguishes ReVi's dual-loop mechanism from these methods.

A2-1. ReVi builds on both methods and cites them: its editing-policy search adopts GEPA's natural-language reflection on execution traces [Agrawal et al., 2026], and its meta level adopts Promptbreeder's idea that the improvement procedure can itself be improved [Fernando et al., 2024]. ReVi goes beyond both in three respects.

(1) ReVi makes the allocation of the search budget learnable. Promptbreeder evolves how to mutate a prompt, and GEPA runs a fixed reflection procedure. Neither lets the system learn where to spend its evaluation budget. In ReVi, the meta policy μ\mu contains diagnosis rules, proposal rules, and budget-allocation rules, and all three can be revised. An accepted meta policy can, for example, favour testing alternative Paper Edits over repeated video generation. This matters for video generation, where one rendered evaluation costs dollars rather than tokens, so where the budget goes changes the outcome as much as how each candidate is written.

(2) ReVi judges an improvement procedure by the whole search it produces. Promptbreeder scores a mutation prompt through the fitness of the single offspring it produces. ReVi evaluates a candidate meta policy over a complete episode: the parent editor, the production traces, the improvement history, and the search budget KK are fixed, the candidate runs a full editor search, and it is scored by the improvement of the resulting editor on held-out novels (Eq. 4). A meta policy is therefore rewarded for making the search more productive, not for one lucky mutation.

(3) ReVi attributes each failure to the stage that caused it. GEPA reflects on each module's own trace, so a misaligned visual is explained at the visual stage. ReVi's artifacts are source-linked across all stages, so the meta policy can trace that symptom to its cause, for example insufficient screen time allocated by the Radio Editor, and patch the allocation rule instead. A per-video failure thus becomes a reusable rule change at the right stage.

AspectPromptbreederGEPAReVi
Editable improvement procedureMutation promptsFixedDiagnosis, proposal, and budget-allocation rules
How the procedure is judgedFitness of its offspring—Improvement produced by a complete, budget-matched search
Failure attributionNonePer moduleEarliest responsible stage, through source links
AcceptanceTournament selectionPareto frontier over instancesFrozen kernel with hard constraints

W2-2. What specific advantage the meta-policy provides.

A2-2. The meta policy roughly doubles the gain of the outer loop at the same cost. In the ten paired runs of Appendix D.3, editing-policy revision raises QQ by +2.26 over Frozen, and adding meta-policy revision raises it to +4.47. The meta-policy term is positive in all ten runs, by 1.67 to 2.60 points, and it improves all three dimensions (+2.22 RF, +1.99 NP, +2.41 AVG). The two conditions spend within about one dollar per run of each other, and the matched-candidate experiment in A3-1 shows that the gain does not come from evaluating more candidates.

AuthorsOpenReview ↗

Response to Reviewer pM8W [3/7]

[W3] The outer-loop experiments do not adequately establish the benefit of the proposed optimization strategy. The two optimization settings evaluate different numbers of candidates (147 versus 195), making it difficult to separate the contribution of the meta-policy from the effect of a larger search budget. A comparison with an existing prompt optimizer under comparable evaluation budgets would provide stronger evidence.

Response: We thank the reviewer for this important comment. The concerns center on whether the meta-policy gain comes from a larger number of candidates, and how ReVi compares with an existing prompt optimizer under a comparable budget. We try to address each concern below.

W3-1. The two optimization settings evaluate different numbers of candidates (147 versus 195), making it difficult to separate the contribution of the meta-policy from the effect of a larger search budget.

A3-1. We first explain where the difference comes from, and then add an experiment with the same number of candidates.

(a) Why the counts differ. The 195 candidates of Editing-Policy + Meta-Policy are 149 editing-policy candidates plus 46 meta-policy candidates, while Editing-Policy evaluates 147 editing-policy candidates. The two conditions therefore test almost the same number of editors; the extra 46 candidates are proposals for the improvement procedure itself. They also cost little. During the search, each candidate is screened at the plan level without rendering video, and after each round the promoted policy is rendered once on each of the six search novels. Both conditions therefore render the same 30 recaps per run, and rendering dominates the cost.

Rounds completed12345
Editing-Policy: total cost per run (USD)35.0169.44102.84137.24172.54
Editing-Policy + Meta-Policy: total cost per run (USD)34.0868.39102.74137.95173.47

As shown in the table, the two conditions differ by at most $1.05 per run at any round, including the cost of all 46 meta-policy candidates.

(b) An experiment with the same number of candidates. To remove the remaining difference, we rerun Editing-Policy with its candidate budget raised to that of the meta-policy condition. Specifically, each of the ten Editing-Policy runs is given the same number of editing-policy candidates as its paired Editing-Policy + Meta-Policy run (195 in total), with the same search procedure, kernel, acceptance rule, and six search novels. The Frozen and Editing-Policy + Meta-Policy results are reused from Appendix D.3.

ConditionCandidates (10 runs)RFNPAVGΔQ vs. Frozen
Frozen091.0588.5682.87—
Editing-Policy (Appendix D.3)14793.2491.0584.97+2.26
Editing-Policy, matched candidates19593.2791.3585.00+2.38
Editing-Policy + Meta-Policy (Appendix D.3)19595.4693.0587.38+4.47

As shown in the table, giving the editing-policy search 48 more candidates raises its gain only to +2.38, well below the +4.47 of the meta-policy condition at the same candidate count. The meta-policy gain therefore does not come from a larger search budget: the extra candidates help little when the improvement procedure stays fixed, while a revised procedure makes the same number of candidates more productive.

W3-2. A comparison with an existing prompt optimizer under comparable evaluation budgets would provide stronger evidence.

A3-2. We agree. We would like to clarify that ReVi's editing-policy search already adopts GEPA-style reflective mutation [Agrawal et al., 2026] (Section 2.2): it reads production traces and critic reports, diagnoses failures in natural language, and proposes revised instructions. The two outer-loop conditions therefore differ from GEPA in a well-defined way.

ComponentGEPAEditing-PolicyEditing-Policy + Meta-Policy (ReVi)
Reflective mutation on traces✓✓✓
Candidate selectionPareto frontier and mergeKernel promotionKernel promotion
Editable improvement procedure✗✗✓

As shown in the table, the matched-candidate comparison in A3-1 changes only the last row, so it measures exactly what a fixed reflective optimizer lacks. GEPA's Pareto selection concerns a different component and is complementary to the meta level; it could serve, for example, as the editor-selection rule inside each meta-policy episode.

Reviewer_u7JGOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

The goal of this paper is to generate a short audiovisual recap directly from a novel within a fixed viewing-time budget while preserving the main storyline and casual connections. The proposed ReVi, uses an inner loop and an outer loop. The inner loop selects and compresses story events, turns them into timed narration, and plans corresponding shots before generating the video. It also checks the output and revises earlier stages when needed. The outer loop learns from production records acorss stories to improve reusable editing rules and the process used to revise those rules, without updating the underlying model parameters. The main contributions are the novel-to-recap task formulation, an editing-first generation pipeline and a self-improvement mechanism that transfers experience across stories.

Critical strengths
  1. Clear writing and organization. The paper is easy to follow. It clearly explains the task, the role of each editing agent, and how the inner and outer loops work together.
  2. Novel framework. ReVi not only revises the narrative structure within an indivisual story, but also learns from experience across stories to improve its novel-to-recap generation process.
Critical weaknesses
  1. The two-stage alternative is not evaluated. The paper motivates direct novel-to-recap generation by discussing the limitations of novel-to-film and film-to-recap, but does not compare against a pipeline combining these two steps. Such a comaprison would help establish the direct approach's advantages in recap quality, production cost and generation time.
  2. Character fidelity to the novel is unclear. The paper mentions visual references and character compatibility checks, but does not explain how the references are created from the novel or verified against its descriptions. Keeping a character's appearance consistent across shots deos not necessarily mean that the appearance matches the source text.
Questions for authors
  1. Can you add a comparsion to the two-stage alternative?
  2. Does the recap help unfamiliar viewers understand the story?
  3. Does meta-policy improvement remain beneficial under the same total computation budget?
8 comments
AuthorsOpenReview ↗

Response to Reviewer u7JG [1/4]

We sincerely thank the reviewer for the careful reading and the constructive comments. We thank the reviewer for recognizing the clear writing and organization, the explicit roles of the editing agents, and the framework that revises the narrative structure within a single story while learning from experience across stories.

[W1] The two-stage alternative is not evaluated. The paper motivates direct novel-to-recap generation by discussing the limitations of novel-to-film and film-to-recap, but does not compare against a pipeline combining these two steps. Such a comparison would help establish the direct approach's advantages in recap quality, production cost and generation time.

Response: We thank the reviewer for this suggestion. We compare the four possible routes to a recap along the attributes that matter for the task: access to the source, recap quality, duration compliance, narration–visual alignment, and production cost. The comparison uses existing results, because the first stage of the two-stage pipeline is already part of our evaluation: ViMax is a novel-to-film system, so the main comparison measures the material that a film-to-recap editor would start from.

A1-1. The four routes differ in when and where the story is compressed.

RouteAccess to the source textRecap fidelityDuration complianceNarration–visual alignmentProduction cost
Novel-to-filmHighHighLowLowHigh
Film-to-recap (on an existing film)LowLowHighMediumMedium
Novel-to-film + film-to-recapHighLowLowLowHigh
ReViHighHighHighHighLow

What each entry rests on is summarised below.

RouteWhy
Novel-to-filmReads the whole novel, so nothing is lost upstream, but the film is authored for its own runtime: ViMax produces 396.2 s for a 180 s target and renders 540 s at $8.76 per novel, with 0/12 outputs within the duration tolerance. Narration is written for dialogue scenes rather than against a narration track.
Film-to-recap (on an existing film)Controls the output length well, since it only cuts existing footage, but it sees only the film: events that the adaptation omitted cannot be recovered, so its fidelity is bounded by the film it starts from. It also depends on an adaptation that must already exist.
Novel-to-film + film-to-recapCombines the two: it has access to the source, yet compresses only after the adaptation has already discarded events, and it pays for the full film before any compression begins (540 s rendered, $8.76 per novel). The second stage must also fit narration to dialogue footage that was not planned for it.
ReViCompresses the story while the source text is still available (Paper Editor with causal closure), allocates narration time by narrative function and verifies it against measured TTS before rendering (Radio Editor), and plans each shot against the narration and its own Beat (AV Planner). It meets the 180 s target in 12/12 outputs at $3.87 per novel, with RF/NP/AVG of 90.68/88.25/82.88.

As shown in the tables, the two-stage route keeps only one advantage of each stage and inherits the weakness of both: it has the source available, but its fidelity is capped by what the first stage preserved, and it pays the full adaptation cost before editing starts. ReVi obtains every attribute at once because the compression decision is made while the source text is still in hand.

AuthorsOpenReview ↗

Response to Reviewer u7JG [2/4]

[W2] Character fidelity to the novel is unclear. The paper mentions visual references and character compatibility checks, but does not explain how the references are created from the novel or verified against its descriptions. Keeping a character's appearance consistent across shots does not necessarily mean that the appearance matches the source text.

Response: We thank the reviewer for this precise observation. The concerns center on how character references are created from the novel, and how they are kept faithful to the source text rather than only consistent across shots. We try to address each concern below.

W2-1. The paper mentions visual references and character compatibility checks, but does not explain how the references are created from the novel or verified against its descriptions.

A2-1. We agree that the manuscript describes this step too briefly. In ReVi, a character's appearance is derived from the novel's own words, and this text-grounded description anchors every generated image.

StageWhat ReVi doesWhat it ensures
Source groundingThe Story Mapper links each event to its supporting source passages. For each character, the appearance description is drawn from the passages that name the character, so the attributes come from the novel's own wording rather than from a synopsis or from another character.Appearance comes from the source text
Reference imageA reference image is generated from that description, and a three-view sheet is built from it: the front view is generated first, and the side and back views use it as the identity reference.One identity across views
Shot bindingEvery generation request carries the character's reference image, and the AV Planner groups shots only when the characters, settings, and state transitions in the group are compatible.Every shot is anchored to the same source-grounded reference, so the appearance cannot drift between shots
Unstated attributesFidelity is defined over the attributes the novel states. A trait the novel leaves open is fixed once by the canonical reference and kept for the whole video.No drift between shots where the novel is silent

W2-2. Keeping a character's appearance consistent across shots does not necessarily mean that the appearance matches the source text.

A2-2. We agree, and in ReVi the two properties are enforced at two different places, which is why consistency does not have to stand in for fidelity.

Fidelity to the source is fixed when the reference is built. The appearance description is written from the source passages that name the character, and the reference image and three-view sheet are generated from that description before any shot is rendered. Every attribute in the reference therefore traces back to a passage in the novel, so an appearance that the text does not support cannot enter the reference in the first place.

Consistency is fixed when the shots are generated. Every generation request carries the character's reference image, and the AV Planner groups shots only when the characters, settings, state transitions, timing, and continuity in the group are compatible. Because the reference is the same object in every request, the appearance in the video is the appearance that was checked against the novel, rather than a chain of per-shot redrawings in which an error can drift.

For characters whose appearance changes during a story, the same description stage carries the change: a state variant inherits the parent character's identity and changes only the specified age or costume state. In the reported experiments, every character uses a single canonical reference, which keeps the character recognizable throughout the recap, and appearance changes are carried by the shot descriptions of the Beats that cover them. Automatic detection of appearance changes is a direction we will discuss as future work.

AuthorsOpenReview ↗

Response to Reviewer u7JG [3/4]

[Q1] Can you add a comparison to the two-stage alternative?

Response: We thank the reviewer for this question. Yes. Please refer to A1-1 for the detailed comparison.

[Q2] Does the recap help unfamiliar viewers understand the story?

Response: We thank the reviewer for this question. Yes, and we test this directly with viewers who do not know the novels.

The existing human study points in this direction: ReVi receives the highest Narrative Pacing rating (4.56 vs. 3.94 for ViMax), and NP measures how clearly a recap develops its story. Those raters, however, could consult the source text, so they judged the recaps as informed readers rather than as new viewers.

Experimental setup.

  1. Viewers. Six new viewers, separate from the raters of the existing study, who report that they have neither read these novels nor watched an adaptation of them.
  2. Materials. Six of the 12 Recap32 test novels (N01, N03, N05, N07, N09, N11), under three methods: MovieAgent, ViMax, and ReVi. For each novel we fix 10 NarrativeQA questions about the main storyline (goals, key events, causes, and outcomes) before any viewing, and grade with the questions' human-written reference answers.
  3. Design. A Latin square over novels and methods: each viewer watches all six novels, each under one method, so that every viewer sees every method twice and each novel–method pair is watched by two different viewers (36 viewings). Since every viewer is exposed to every method, differences between viewers cannot explain the result.
  4. Procedure. After each recap, the viewer answers the 10 questions in free text without pausing or rewinding, and rates overall understanding on a 1–5 scale. A grader who does not know which method produced the recap scores each answer as correct, partial, or incorrect.

We use human viewers rather than a language model for this question. These are public-domain novels, so a model may already know the plot from pretraining, and a recap could prompt recall rather than convey the story; model accuracy therefore cannot separate what the video taught from what the model already knew.

6 Recap32 test novels, 6 viewers, 10 questions per novelQA accuracy (%)Understanding (1–5)
MovieAgent50.833.17
ViMax55.833.33
ReVi70.003.75

As shown in the table, viewers who had never read the novels answer 1.42 more questions correctly after watching a ReVi recap than after a ViMax recap, and rate their understanding 0.42 points higher. The largest gains are on questions about goals, which is the information that causal editing is designed to preserve.

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

This paper formulates the task of generating a concise audiovisual recap directly from a novel under a fixed duration budget. It proposes ReVi, a dual-loop multi-agent framework based on an “edit first, generate second” workflow. The inner loop constructs a source-linked story graph, selects and compresses events through a Paper Editor, allocates narration time through a Radio Editor, and plans temporally aligned visual beats through an Audio-Visual Planner. A repair mechanism routes detected errors back to the earliest responsible stage. The outer loop uses production traces to revise both the reusable editing policy and the meta-policy that governs policy revision. On 15 FilmEval narratives and 12 longer Recap32 novels, ReVi obtains higher automatic recap-fidelity, narrative-pacing, and audio-visual-grounding scores than several baselines, while satisfying the target duration in all reported cases. A shared-narration control also shows a ReVi advantage in downstream video realization. Human ratings on six FilmEval sources favor ReVi, although the human study is small. Additional ablations suggest that causal editing, narration-time allocation, beat planning, and cross-stage repair each contribute to different aspects of performance.

Critical strengths
  1. A coherent decomposition of the novel-to-recap problem. ReVi treats event selection, narration timing, and visual planning as coupled but separable decisions. The source-linked intermediate representations and cross-stage repair mechanism provide a sensible way to diagnose whether a failure originates in content selection, timing, or visual realization.

  2. Strong empirical gains on both duration compliance and measured recap quality. Across FilmEval and Recap32, ReVi consistently improves the reported quality metrics and meets the duration tolerance for all outputs. The shared-narration control is particularly useful because it shows that the advantage is not only due to ReVi writing a different narration script.

  3. A thoughtful attempt to evaluate recursive policy improvement. The paper distinguishes revisions to the task policy from revisions to the policy-improvement procedure itself, and uses paired searches and a frozen acceptance kernel rather than claiming unrestricted self-improvement.

Critical weaknesses
  1. The main baseline comparison is confounded by unequal duration control and repair budgets. ReVi explicitly measures TTS duration, allocates narration time before rendering, performs cross-stage repairs, and allows regeneration. The baselines receive the target duration only in their prompts, while their original pipelines do not enforce the duration constraint. Consequently, every baseline has a 0% duration-pass rate, whereas ReVi has 100%. This makes the large quality and duration gap partly a comparison between a constrained, repaired system and unconstrained baselines. A stronger comparison would give all methods the same duration-enforcement and post-generation repair framework, or evaluate all outputs after a common duration-constrained procedure.

  2. The evaluation relies heavily on correlated automatic judgments and small, incomplete human validation. The three Gemini judgments are repeated evaluations of the same videos rather than independent generations, and the per-segment judgments needed to recompute RF and AVG are unavailable. The human study covers only six FilmEval sources and has moderate inter-rater agreement, especially for recap fidelity and pacing. No human evaluation is reported for the 12 Recap32 test novels. Thus, the reported 8–22 point improvements are promising but should be interpreted as improvements under the paper’s automatic evaluator rather than fully established human-perceived gains.

  3. The recursive-improvement evidence has limited independent generalization. The repeated-search analysis uses six fixed novels across ten runs and multiple revision rounds, with the same source identities reused across conditions and rounds. This provides useful paired evidence, but it does not establish robustness across a broad population of novels or independent generations. The paper should clarify whether any search decisions, candidate promotion, or policy selection had access to outputs from the reported “test” novels.

Questions for authors
  1. Can the authors rerun the main baselines with the same target-duration enforcement, regeneration budget, and cross-stage repair opportunities as ReVi, or otherwise report quality after a common duration-constrained postprocessing procedure? This is necessary to separate the contribution of the editorial framework from the advantage of explicit runtime control.

  2. Please clarify the exact data flow in the outer-loop experiments. Were any outputs, scores, or candidate-selection decisions from the six novels used in the repeated-search analysis incorporated into policy promotion or final policy selection? An independent evaluation on unseen novels would strengthen the recursive-improvement claim.

  3. Can the authors release per-video automatic judgments, source references, and human ratings, or provide an independent human evaluation on the Recap32 test set? The current aggregate scores and small six-source human study make it difficult to assess evaluator reliability and generalization.

8 comments
AuthorsOpenReview ↗

Response to Reviewer tgzh [1/5]

We sincerely thank the reviewer for the careful reading and the constructive comments. We thank the reviewer for recognizing the coherent decomposition of the novel-to-recap problem, the strong empirical gains in both duration compliance and recap quality, the usefulness of the shared-narration control, and our effort to distinguish revisions of the task policy from revisions of the policy-improvement procedure.

[W1] The main baseline comparison is confounded by unequal duration control and repair budgets. ReVi explicitly measures TTS duration, allocates narration time before rendering, performs cross-stage repairs, and allows regeneration. The baselines receive the target duration only in their prompts, while their original pipelines do not enforce the duration constraint. Consequently, every baseline has a 0% duration-pass rate, whereas ReVi has 100%. This makes the large quality and duration gap partly a comparison between a constrained, repaired system and unconstrained baselines. A stronger comparison would give all methods the same duration-enforcement and post-generation repair framework, or evaluate all outputs after a common duration-constrained procedure.

Response: We thank the reviewer for this important comment. The concerns center on whether the baselines receive an equal opportunity to control duration, and whether the quality gap survives when they do. We try to address each concern below.

W1-1. The baselines receive the target duration only in their prompts, while their original pipelines do not enforce the duration constraint. Consequently, every baseline has a 0% duration-pass rate, whereas ReVi has 100%.

A1-1. We agree that the baselines should also be given explicit duration control. We therefore rerun the two strongest baselines, ViMax and MovieAgent, with a duration-control prompt at every planning stage. We choose these two because ViMax is the strongest baseline on AVG and MovieAgent the strongest baseline on RF, on both datasets. ReVi is unchanged and reuses its main-comparison results.

Experimental setup.

  1. Test sets. The same test sets as the main comparison: 15 FilmEval sources (90/120/150 s) and 12 Recap32 test novels (180 s), with the same generation models and the same judge.
  2. Duration-control prompt. The target playback length is written into the prompt of every planning stage of each baseline (script, narration, and storyboard), not only into the top-level task prompt. The narration prompt states the corresponding word budget at the speaking rate of the shared TTS voice, and the storyboard prompt requires the shot durations to sum to the target.
  3. Everything else unchanged. Each baseline keeps its own planning modules and rendering calls.
DatasetMethodRFNPAVG
FilmEvalMovieAgent, original prompt83.7278.6174.73
FilmEvalMovieAgent, duration-control prompt82.4677.6374.67
FilmEvalViMax, original prompt82.0978.4478.66
FilmEvalViMax, duration-control prompt80.9277.1179.00
FilmEvalReVi (Table 1)91.5890.2088.67
Recap32MovieAgent, original prompt80.9977.5072.30
Recap32MovieAgent, duration-control prompt79.7576.3872.92
Recap32ViMax, original prompt78.3076.3574.07
Recap32ViMax, duration-control prompt76.8075.1674.17
Recap32ReVi (Table 1)90.6888.2582.88

As shown in the table, the duration-control prompt brings the baseline outputs close to the target (mean length 179.9 s on Recap32, against 396.2 s for ViMax and 524.4 s for MovieAgent with the original prompt), but it does not close the quality gap. ReVi leads the duration-controlled ViMax by 13.88 / 13.09 / 8.72 RF/NP/AVG points and the duration-controlled MovieAgent by 10.93 / 11.87 / 9.97 on Recap32, with the same pattern on FilmEval. When a baseline is forced into the target length, it must drop content without knowing which events the retained story depends on, so fitting the length alone does not raise fidelity or pacing. ReVi's advantage therefore comes from its editorial decisions: which events to keep by their causal role, and how much narration time each one receives.

AuthorsOpenReview ↗

Response to Reviewer tgzh [2/5]

W1-2. This makes the large quality and duration gap partly a comparison between a constrained, repaired system and unconstrained baselines. A stronger comparison would give all methods the same duration-enforcement and post-generation repair framework, or evaluate all outputs after a common duration-constrained procedure.

A1-2. We respectfully clarify that the manuscript already evaluates all methods under a common duration-constrained procedure, and measures the contribution of repair separately.

(a) A common duration-constrained procedure (Appendix D.1). In the shared-narration control, every method receives the same human-edited narration script and the same audio file for each source, so spoken content and narration timing are identical across methods. The baselines' duration-pass rate rises to 46.7–86.7% on FilmEval and 41.7–91.7% on Recap32.

ReVi − ViMax under shared narrationΔRFΔNPΔAVG
FilmEval, all 15 sources+6.47+6.12+6.18
FilmEval, 13 sources where both pass the 5% tolerance+6.63+6.00+6.46
Recap32, all 12 novels+6.02+6.72+6.72
Recap32, 11 novels where both pass the 5% tolerance+6.19+6.68+6.86

As shown in the table, ReVi keeps a lead of about six points on every dimension when narration and timing are identical for all methods, including on the subset where both methods satisfy the duration tolerance.

(b) Repair and regeneration. We do not transplant ReVi's cross-stage repair into the baselines, because repair routing traces each failure through ReVi's source-linked artifacts to the earliest responsible stage, and the baselines' pipelines have no such artifacts to route through. Its contribution is instead measured inside ReVi: removing repair routing lowers RF/NP/AVG by 4.64/3.10/3.17 points (Appendix D.2), which is smaller than the gap to the baselines. Regeneration does not favour ReVi either. The baselines already generate proportionally more discarded or regenerated footage than ReVi: the ratio of generated to final video is 1.362 for ViMax against 1.214 for ReVi on FilmEval, and 1.358 against 1.172 on Recap32 (Appendix D.4).

AuthorsOpenReview ↗

Response to Reviewer tgzh [3/5]

[W2] The evaluation relies heavily on correlated automatic judgments and small, incomplete human validation. The three Gemini judgments are repeated evaluations of the same videos rather than independent generations, and the per-segment judgments needed to recompute RF and AVG are unavailable. The human study covers only six FilmEval sources and has moderate inter-rater agreement, especially for recap fidelity and pacing. No human evaluation is reported for the 12 Recap32 test novels. Thus, the reported 8–22 point improvements are promising but should be interpreted as improvements under the paper's automatic evaluator rather than fully established human-perceived gains.

Response: We thank the reviewer for this careful assessment. The concerns center on the transparency of the automatic judgments and the coverage of the human validation. We try to address each concern below.

W2-1. The three Gemini judgments are repeated evaluations of the same videos rather than independent generations, and the per-segment judgments needed to recompute RF and AVG are unavailable.

A2-1. We agree that the three judgments reduce judge variance but do not remove judge bias; the manuscript accordingly describes them as repeated judgments rather than independent generations (Appendix C). To make every reported FilmEval score auditable, we release the per-video automatic judgments of all 105 FilmEval videos in the main comparison (15 sources × 7 methods) through the anonymous repository.

Field groupContent
IdentifiersSource ID (C1–C15), difficulty tier, target duration, method
Automatic judgmentsRF, NP, and AVG of each video (median of the three judge calls)
DurationFinal length, duration error, and pass flag at the 5% tolerance
CostGenerated length and generation cost of each video

W2-2. The human study covers only six FilmEval sources and has moderate inter-rater agreement, especially for recap fidelity and pacing. No human evaluation is reported for the 12 Recap32 test novels.

A2-2. We agree that the human evidence should also cover the book-length novels. We first note that the existing raters agree closely in absolute terms: on the 42 FilmEval videos, the three raters' scores lie within one point of each other on 42/42 videos for RF, 39/42 for NP, and 41/42 for AVG. Human raters prefer ReVi over ViMax by 0.28/0.61/0.50 points on the 0–5 scale, and automatic and human scores agree with pooled correlations of 0.831/0.881/0.899. We further add a human evaluation on the Recap32 test set.

Experimental setup.

  1. Videos. All 12 Recap32 test novels × 3 methods (MovieAgent, ViMax, ReVi), giving 36 videos. These are the main-comparison outputs, so the human and automatic evaluations judge the same videos.
  2. Raters and protocol. Three raters score every video blind and in randomized order, on the same 0–5 RF, NP, and AVG scales and with the same written rubric as the FilmEval study.
  3. References. Since a rater cannot read a 70k–190k-word novel within a session, each rater receives the human-written NarrativeQA plot summary of the novel together with our shared event reference.
MethodRFNPAVG
MovieAgent4.063.893.47
ViMax3.923.753.83
ReVi4.394.444.33

As shown in the table, human raters prefer ReVi over ViMax by 0.47 / 0.69 / 0.50 points on the book-length novels, and the ranking of the three methods matches the automatic ranking on every dimension. The three raters' scores lie within one point of each other on 36 of the 36 videos. These results indicate that ReVi's gains are perceived by human viewers on long novels as well as on FilmEval, rather than only under the automatic evaluator.

Comments on the paper

1 comment

Hi Reviewers,

The authors posted their replies along with the new results. Please you please review them and point out any remaining concerns that you may have?

Thank you

AC

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.