acceptodds
Under review as a conference paper at ICLR 2027

CineForge: Self-Improving Agents for Long-Horizon Video Generation

Abstract

Long-horizon story-driven video generation requires coordinating narrative decomposition, state tracking, shot design, prompting, rendering, and revision across interdependent scenes. Adaptive video systems primarily refine requests or reusable skills, leaving recurring failures disconnected from persistent, stage-targeted improvements across stories. We introduce CineForge, a self-evolving video-production agent framework coupling CineForge-Produce for generation with CineForge-Evolve for cross-story policy evolution. CineForge-Produce structures stories into typed narrative, character, spatial, and cinematic states to coordinate asset and clip generation, recording canonical production trajectories. CineForge-Evolve applies Case-to-Pattern-to-Policy Evolution (CPPE) to review trajectory evidence, consolidate recurrent findings into bounded stage-local patches, and deploy validated updates through structural replay and confidence-controlled paired evaluation. To measure complete story realization, CineScope combines 100-script CineScope-Data with human-aligned, multiscale CineScope-Metric covering causal state, directorial orchestration, pacing and resource allocation, and character arc. Across CineScope-Data and two public benchmarks, the evolved policy improves CineScope-Metric from 4.02 to 4.38, outperforms three long-video baselines with consistent gains under ScriptAgent, and reduces review LLM calls by 37.0% on new stories. Results establish production trajectories as actionable experience for cumulative improvement across long-form storytelling.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Recursive Self-Improvement, Long-Horizon Video Generation, Agentic Optimization

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_xaZ5OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper presents CineForge, a self-improving agent framework for long-horizon story-driven video generation. The framework consists of CineForge-Produce, which organizes video production into structured stages and records production trajectories, and CineForge-Evolve, which uses Case-to-Pattern-to-Policy Evolution (CPPE) to identify recurring failures and update persistent production policies. Experiments show that CineForge improves CineScope scores after three evolution rounds, outperforming adapted GenMAC and VideoWeaver baselines. The results also suggest improvements on unseen stories and reduced review workloads.

Critical strengths
  1. An interesting and practical research problem. The paper addresses an important limitation of existing agentic video generation systems: the difficulty of transferring lessons from previous production failures to future stories. The distinction between instance-level repair and persistent production-policy improvement is meaningful.

  2. A well-structured production and evolution framework. The separation between CineForge-Produce and CineForge-Evolve is clear. The use of production trajectories, cross-story failure patterns, bounded policy patches, and replay-based validation provides a reasonably systematic framework for self-improvement.

  3. Encouraging experimental results. The proposed method achieves improvements over several video generation and optimization baselines. The evaluation includes human judgments, external metrics, unseen-story comparisons, and component ablations. The CineScope benchmark also provides a useful perspective on story-level video quality.

Critical weaknesses
  1. Limited methodological novelty. Ideas are closely related to existing work. Although the paper discusses these connections, it does not sufficiently establish which components are fundamentally new. The novelty appears to lie mainly in the integration of existing mechanisms into a video production pipeline.

  2. A gap between the theoretical guarantees and their empirical verification. The paper presents confidence-controlled replay and structural improvement guarantees as important contributions. However, Appendix B.9 acknowledges that the reported experiments do not provide the complete per-candidate numerical audit needed to verify every deployed update.

  3. Potential fairness issues in the video generation comparisons. Although the methods share the same underlying models and generation budgets, their realized output lengths differ substantially. In Table 7, CineForge-Produce generates videos with a median duration of 12.0 minutes, compared with 3.52 minutes for ViMax and 2.18 minutes for AniMaker. Such differences may affect narrative coverage, character development, and other story-level evaluation dimensions. It is therefore difficult to attribute all improvements solely to the proposed production framework.

  4. Limited evaluation of generalization and evaluator reliability. CineScope-Metric achieves reasonable agreement with human judgments, but its held-out alignment evaluation uses only 15 distinct stories, and some individual dimensions show moderate correlations. The proposed method is evaluated primarily on a relatively small fixed test set, while the evolution experiments do not report uncertainty across independently learned policies.

Questions for authors
  1. Could the authors further clarify the main methodological differences between CPPE and existing trajectory-driven agent or workflow optimization approaches? A comparison with a general-purpose policy optimization baseline would also be helpful.

  2. Could the authors provide additional details on the numerical verification of deployed policy updates, including acceptance margins and regression-test results? It would also be helpful to clarify how the theoretical guarantees relate to the actual sampling protocol.

  3. Could the authors discuss how differences in generated video duration might influence the comparisons? A duration-controlled evaluation would help strengthen the conclusions.

  4. Could the authors provide further evidence of CineScope-Metric's reliability on a larger held-out set? Additional independent evolution runs would also help assess the stability and generalization of the proposed method.

  5. Could the authors provide a more comprehensive analysis of the computational costs associated with policy evolution, including candidate validation and replay? This would help clarify the practical efficiency benefits of the framework.

Reviewer_oW81OpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper presents CineForge, an agentic system for long-form, story-driven video generation that improves its production policy across stories. CineForge-Produce decomposes a story into narrative atoms, episode/scene/shot plans, and character/spatial ledgers, generates and concatenates clips, and logs a full production trajectory. CineForge-Evolve (CPPE) localizes failures to the earliest responsible stage, groups failures that recur across stories into patterns, compiles them into bounded stage-local text-policy patches, and admits at most one patch per round through replay-based confidence gates. The paper also introduces CineScope, a 100-script benchmark with a VLM-based, human-aligned metric. Experiment results show the proposed method outperforms adapted GenMAC and VideoWeaver optimizers.

Critical strengths
  • All data splits are disjoint, and the outcome evaluator is frozen and kept out of the optimization loop.
  • On held-out stories, the human score gain (+0.32) matches the CineScope gain (+0.36).
Critical weaknesses
  • No gate statistics or deployed patches are reported, and App. B.9 concedes that the per-candidate audit was not done.
  • The only patch shown (App. A.2) is a generic prompt rule, yet there is no comparison with a Round 0 policy that includes such rules from the start, or with open-source prompt optimizers (e.g., GEPA).
Questions for authors
  1. Could the authors report the admitted patches and the gate statistics for each round?
  2. Could the authors compare against a Round 0 policy with hand-written rules and an open-source prompt optimizer (e.g., GEPA) on the same backbone?
Reviewer_xYtLOpenReview ↗
Rating
3
Confidence
3
Informativeness
3
Presentation
2
Soundness
2
Read review
Paper summary

This paper introduces CineForge, a self-improving agent framework for long-horizon story-driven video generation. The system has two components. CineForge-Produce decomposes a story into narrative atoms, episodes, scenes, shots, character states, spatial states, assets, prompts, and generated clips, while recording a canonical production trajectory. CineForge-Evolve applies a Case-to-Pattern-to-Policy Evolution procedure that identifies recurrent failures across productions, localizes them to production stages, compiles bounded stage-local policy patches, and deploys at most one patch per round after structural replay and paired stochastic evaluation. The paper also introduces CineScope, consisting of 100 long-form scripts and a frozen multiscale evaluation metric with twenty criteria grouped into Causal State, Directorial Orchestration, Pacing and Resource Allocation, and Character Arc. The metric is calibrated against human ratings and supplemented with ScriptAgent and VBench-Long evaluations. Across a held-out 45-story system-test split, the evolved system improves CineScope-Metric from 4.02 to 4.38 and improves ScriptAgent from 3.91 to 4.10. Human evaluation also reports an improvement from 3.86 to 4.18 over evolution rounds. The evolved policy reduces review calls and detected issues under a frozen audit configuration. Ablations suggest that trajectory-grounded localization, cross-story recurrence, typed patches, and fresh replay screening all contribute to the observed gains.

Critical strengths
  1. A coherent framework for persistent, cross-story improvement. CineForge distinguishes instance-level repair from policy-level evolution and records sufficient intermediate state to associate failures with production stages. The Case-to-Pattern-to-Policy design, recurrence requirement, typed patches, and rollback metadata form a reasonably principled approach to improving a multi-stage generative workflow.

  2. A relatively strong evaluation protocol for a difficult task. The authors use disjoint collection, alignment, replay-admission, and system-test splits; repeated generations; frozen outcome evaluators; human validation; and ablations. The improvements from Round 0 to Round 3 are visible on CineScope, ScriptAgent, and human ratings, which makes the evidence more convincing than a single automatic score.

Critical weaknesses
  1. The contribution is difficult to separate from extensive prompt engineering and system-specific infrastructure. CineForge-Evolve appears to improve a text-space production policy using LLM-based diagnosis, patch proposal, and rule insertion. The paper provides a sophisticated interface and many safeguards, but it is less clear which component is the essential algorithmic contribution and which gains come from additional engineering, review calls, or accumulated prompt/context information. The comparison to GenMAC and VideoWeaver is useful, but these are adapted optimizer baselines and do not fully establish that the particular CPPE abstraction is superior to other general policy-search or memory mechanisms under matched optimization cost.

  2. CineScope-Metric is useful but remains an imperfect foundation for the headline claims. The overall held-out correlation with human ratings is reasonable, but the dimension-level correlations are substantially lower, especially for Directorial Orchestration. The metric is itself a VLM-based evaluator with eligibility, abstention, temporal weighting, and prompt calibration choices. Small calibration and held-out sets, together with extensive evaluator design degrees of freedom, leave room for metric-specific effects. Human evaluation of the evolution gains helps, but the main system comparisons and most ablations still rely on the automatic metric.

  3. The external validity of the results is limited. CineScope-Data contains only 100 scripts, while the final system-test split contains 45 stories across three sources. The experiments use a fixed text backbone and Seedance 2.0 video generator, so it is unclear whether the learned policy patches transfer across video backbones, providers, or substantially different production interfaces. The benchmark corpus also draws from novels, stories, and screenplays whose detailed provenance and licensing information are withheld during review, which may complicate independent reproduction.

  4. Efficiency claims are narrower than the framing may suggest. The reported reduction in review calls and detected issues is measured under a frozen audit configuration and excludes generation, candidate compilation, replay admission, optimization, and evaluation costs. The paper explicitly acknowledges this limitation, but the headline “self-improving” and workload-reduction claims should be framed more carefully unless total cost, latency, and first-pass yield are also reported.

Questions for authors
  1. Can the authors provide a clearer accounting of total optimization cost for CineForge-Evolve versus GenMAC and VideoWeaver, including trajectory review, candidate generation, replay admission, video generations, and evaluator calls? How do the methods compare under equal total compute or monetary budget?

  2. Which individual design components are responsible for the largest gains? In particular, can the authors report ablations separating trajectory recording, LLM-based localization, cross-story memory, typed patch schemas, and replay gating while keeping the number of review and generation calls fixed?

  3. How robust are the learned policy patches across different video backbones or providers? Does a patch learned with Seedance 2.0 transfer to another generator, or is the policy tightly coupled to the specific backend and prompt interface?

  4. The CineScope-Metric has relatively low dimension-level correlation for Directorial Orchestration. How sensitive are the reported Round 0 to Round 3 gains to alternative temporal weights, eligibility thresholds, prompt variants, or an independently designed evaluator?

  5. Can the authors release the complete CineScope script manifest, task-package format, policy/trajectory schemas, and baseline adapters, subject to licensing constraints? If source scripts cannot be redistributed, what exact artifacts will be available for independent replication?

  6. The paper reports human improvement across rounds, but the confidence intervals and paired statistical tests for the Round 3 minus Round 0 human gain should be made explicit. Is the improvement statistically significant after accounting for story, evaluator, and generation-replicate effects?

Comments on the paper

1 comment

Thank you very much to the authors for the detailed response. The additional clarifications and results regarding computational cost, stability across independent runs, and the extended validation of CineScope have alleviated my concerns to some extent. However, I still have reservations regarding the overall soundness of the work and the degree of human verification in the manuscript, as there remain several instances of overly defensive or self-contradictory phrasing. Below are a few representative examples:

  • Introduction (Lines 97–100): Near the end of the Introduction, the paper abruptly states, "These workload reductions under a frozen audit configuration establish neither lower end-to-end cost nor higher first-pass yield." This disclaimer appears rather abruptly and feels disconnected from the surrounding context.
  • Main text (Lines 370–372): The text states that "Unsupported dimensions are reported as N/A; missing supported scores are not replaced by zero or another dimension," yet all six VBench-Long dimensions reported in Table 2(c) and Table 3(c) contain complete numerical values without any N/A entries.
  • Main text (Lines 411–412) vs. Appendix B.1.3 (Lines 1230–1231): Lines 411–412 report human overall scores as "3.86 [3.80,3.92], 4.059, 4.10 [4.04,4.16], and 4.181 for Rounds 0–3, respectively (available 95% CIs in brackets)," where Rounds 1 and 3 unexpectedly switch to three decimal places and omit confidence intervals. However, Appendix B.1.3 (Lines 1230–1231) states that "All four checkpoints, generation replicates, and associated ratings of a sampled story remain together in each resample." Since all four checkpoints are jointly included within each bootstrap resample, it is unclear why 95% CIs would be available only for Rounds 0 and 2, but not for Rounds 1 and 3.
  • Appendix B.9 (Lines 1622–1627): The appendix notes, "The aggregate quality and workload results reported in this paper do not establish this complete per-candidate numerical audit ... without asserting that every deployed update has been numerically certified under all of the above conditions." Could the authors clarify whether the numerical gates in Algorithm 3 were in fact fully executed for the deployed updates in the reported experiments?
  • Appendix B.3 (Lines 1426–1428): The text writes, "The evaluation report must state valid counts and exclusions alongside the scores; when coverage differs across policies or dimensions, the affected counts and rates must be disclosed." Using prescriptive phrasing (must) rather than a descriptive statement of what was conducted reads somewhat like unedited prompt instructions or internal guidelines. Similar prescriptive patterns appear on Page 30, Line 1617 ("must be computed from... Issue counts cannot be substituted...") and Page 31, Lines 1667–1673 (a sequence of five consecutive "X [claims] require Y" statements).
  • Appendix B.10 (Page 31, Line 1662): The text refers to an unintroduced formula: "The earlier expression [expression] is a special case requiring no new target errors, no removed non-target errors, and unchanged unknown mass." However, this expression does not appear anywhere in the preceding 30 pages. Could the authors clarify the origin of this formula?

The examples above are illustrative rather than exhaustive. Taken together, they leave me with the impression that human verification and end-to-end consistency checking of the manuscript remain somewhat limited, so I will maintain my current score.

Although I do not find the paper ready for publication at ICLR in its current form, I sincerely appreciate the efforts in the rebuttal, and encourage the authors to thoroughly audit the full manuscript and further strengthen the rigor of both the methodology and its implementation. I believe this work holds genuine potential to make a meaningful contribution to long-horizon video generation in the future.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.