acceptodds
Under review as a conference paper at ICLR 2027

EvoReel: Evolving Harness Engineering for Long-Horizon Video Generation

Abstract

Long-horizon video generation turns a brief into a minute-scale, multi-shot film. Since video models typically render clips under 30 seconds, it relies on agentic harnesses that must write a coherent script and render it with characters and places consistent across cuts. Existing harnesses, however, remain static. Across films, nothing learned carries over without fine-tuning or curation, and within a film, every shot is conditioned on references fixed in advance, although each shot needs different characters, places and earlier frames. To address these shortcomings, we propose EvoReel, an evolving harness for long-horizon video generation that learns from its own productions without updating any model. Specifically, EvoReel writes and renders every shot from memory and checks it with a self-judge. Across films, it distills its runs on training films into an experience memory of the revisions that made shots more specific and the content that drafts tended to lose. Within a film, role-specific agents write the screenplay from story to scene to shot over a structured story memory, with the shot writer consulting that experience as it first drafts each scene, and a self-judge scores each draft on professional screenwriting dimensions and sends weak shots back for revision. Rendering follows the same pattern: to keep characters and places consistent across shots, a multimodal visual memory supplies each shot with the reference images and earlier frames it calls for, and a second self-judge checks each rendered shot against its script and re-renders those that fall short. With these designs, EvoReel outperforms the strongest of eight baselines on ViStoryBench and held-out MovieBench by 28.6% and 84.1% in script composite and by up to 20.2% in how much of the brief its films show, and ties for the highest character consistency under the proprietary renderer. We also introduce ReelBench, 104 original premises across 13 genres with checkable constraints, which tests agents beyond shot lists adapted from existing productions. Code and demos are available at https://anonymous.4open.science/r/EvoReel-D2EF.

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

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_MJizOpenReview ↗
Rating
2(+1)
Confidence
4
Informativeness
1
Presentation
1
Soundness
2
Read review
Paper summary

This paper proposes EvoReel, a training-free agentic harness for long-horizon video generation. All models stay frozen. The harness writes the script with role-specific agents over a story memory, renders each shot with references from a visual memory, and uses self-judges to revise weak scripts and re-render failed clips. It also builds an experience memory from past runs to guide new scripts.

Critical strengths
  • The authors put clear effort into the experimental design and the benchmark construction. The overall presentation of the experiments is fairly standard and well organized.
  • The method is close to plug-and-play. It is training-free and works with many kinds of models, both open-source and proprietary.
Critical weaknesses
  • The human evaluation is limited. The study in Appendix A is small and preliminary, with three annotators, and one of them is an author. It is also not used to validate the automatic metrics. For a generative model, human judgment is a key source of evidence. Quantitative metrics and AI judges alone are not enough to show the real quality of the results.
  • The "evolving" concept is overstated. In essence, the method is a memory retrieval process. I do not see the harness itself evolving. This is quite different from AutoHarness or Evo-Bench, where the harness optimizes itself.
  • The overall presentation is unusual. The text reads as if it was heavily polished by AI. The main contributions do not stand out. The paper lists many components, but most of them feel like simple LLM-based designs. The novelty is limited.
  • Figure 3 does not meet the usual standard of the venue. The lines are not aligned, and several labels and symbols are unclear. The fonts and spacing are inconsistent.
  • I am not sure this paper fits ICLR. It also does not seem to fit a CV venue. An ACL-related venue may be a better match.
Questions for authors
  • The anonymous link in the paper does not work. At the time of my review, it showed nothing. This is a video generation paper. Without any demo, even a cherry-picked one, it is hard for me to trust the actual visual quality.
  • As a suggestion, the authors may consider extending the method to audio-video generation, for example with open-source models such as MiniMax-H3. Pure video output has limited expressiveness, and an audio-video version could make the system more practical.
  • Which results use Veo 3.1 and which use LTX-2.3? Could the authors also report the cost of generating one film, including the number of model calls, the token usage, the extra renders from the video self-judge, and the total cost in dollars?
Reviewer_6qviOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper presents EvoReel, an agentic harness for long-horizon, multi-shot video generation with frozen language, vision-language, image, and video models. The main premise is that existing video-generation harnesses are largely static at two levels: they do not retain useful experience across films without model fine-tuning or manually curated knowledge, and they typically rely on references fixed before rendering rather than adapting to what has actually been generated within a film. EvoReel addresses these issues through three interconnected components. First, a script harness decomposes screenplay generation among role-specific agents and maintains a structured story memory, while a script self-judge localizes weaknesses and restricts revisions to predefined writable fields. Second, a video harness maintains a multimodal visual memory that retrieves character, location, and previously rendered visual references for each shot, writes new renders back into memory, and uses a clause-level video self-judge to selectively re-render failed shots. Third, a training harness distills script-generation trajectories on training films into an experience memory containing specificity-improving examples and omission notes, which are retrieved when new scenes are first drafted. The paper evaluates EvoReel on ViStoryBench, MovieBench, and a newly introduced ReelBench, using both proprietary and open-weight video renderers. Against eight baselines, EvoReel achieves the best script-composite score on all three datasets and improves the fraction of briefed content shown on screen on the two shot-list benchmarks. The paper also provides component analyses for the script self-judge, experience retrieval, and video self-judge.

Critical strengths
  1. Practical and renderer-agnostic system design. EvoReel improves long-horizon video generation at the orchestration level without modifying the underlying models. Its structured story memory, dynamic visual references, and separate script/video verification are practically useful, while compatibility with frozen and proprietary renderers improves applicability.

  2. Broad empirical evaluation. The paper compares against eight baselines on three benchmarks and evaluates both script quality and rendered videos. It also includes ablations and component analyses, providing reasonably comprehensive evidence for the effectiveness of the full system.

Critical weaknesses
  1. The novelty of the “evolving harness” framing is not fully convincing. Many components, including experiential memory, hierarchical agents, retrieval-augmented prompting, dynamic memory, and critique-and-repair, have substantial precedent. More importantly, the across-film “evolution” appears to rely on an experience memory mined offline from a fixed training split, rather than continuously incorporating newly generated films. This makes the method closer to offline trajectory distillation plus retrieval than true continual evolution, and the paper could more clearly distinguish its technical novelty from existing memory-based agent systems.

  2. The ablations do not fully support the claimed importance of the core mechanisms. On ReelBench, much of the improvement attributed to the training harness appears to come from the first-draft visual-treatment instruction itself, suggesting that a relatively simple prompting choice may account for a substantial part of the gain rather than experience retrieval. Similarly, the video self-judge provides only limited improvements at the clip and film levels. These results make it difficult to determine how much of EvoReel’s advantage comes from its proposed memory and evolution mechanisms versus the combined effect of several engineering choices.

  3. The contribution of the visual-memory mechanism is not sufficiently isolated. The full system combines retrieval, an LLM selector, a deterministic guard, reference composition, and write-back into memory, but the paper does not clearly quantify how much each component contributes to cross-shot consistency. This is particularly important because the paper positions adaptive visual memory as a central part of the method, while the conclusion explicitly leaves isolation of its contribution to future work.

  4. Some of the paper’s contribution claims appear broader than what is directly supported by the ablations. The paper motivates several distinct mechanisms, including structured story memory, dynamic visual memory with write-back, constrained script revision, experience retrieval, and localized self-judging. However, only a subset of these components is individually evaluated. In particular, the contribution of dynamic visual memory to cross-shot consistency is not directly isolated, and several script-harness design choices are evaluated only as part of the full system. As a result, the experiments demonstrate the effectiveness of the overall pipeline more clearly than they validate the specific claims associated with each proposed mechanism.

Questions for authors
  1. Meaning of “evolving.” The across-film experience memory appears to be constructed offline from a fixed training split and then retrieved during inference. Is this memory updated with newly generated films during deployment, or does it remain fixed after construction? If it remains fixed, could the authors explain why this should be viewed as an evolving harness rather than offline trajectory distillation with retrieval?

  2. Contribution of experience retrieval. Could the authors provide a cleaner ablation separating the effects of the first-draft visual-treatment instruction from experience retrieval, ideally comparing instruction-only, retrieval-only, and their combination under the same inference budget? This would help determine whether the gains attributed to the training harness genuinely arise from accumulated experience.

  3. Role of the visual-memory mechanism. Could the authors isolate the contributions of the main components in the visual-memory pipeline, such as dynamic write-back, the LLM selector, and the deterministic guard? In particular, comparisons against fixed references or previous-frame-only conditioning would help establish whether the proposed evolving visual memory itself is responsible for improved cross-shot consistency.

  4. Contribution of individual harness components. Several mechanisms are motivated as solutions to distinct failure modes, including structured story memory, constrained revision, context-isolated judging, and localized self-judging. Could the authors provide additional component-level ablations, or indicate which of these mechanisms are essential to the observed gains? At present, it is difficult to distinguish the contribution of individual design choices from the aggregate effect of the full pipeline.

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

This paper proposes EvoReel, an evolving agentic framework for long-horizon video generation that improves script writing and video rendering without updating model parameters. The method introduces an experience memory to transfer knowledge across films, structured story and visual memories to maintain consistency within films, and self-judges to verify and refine scripts and generated videos. Experiments on ViStoryBench, MovieBench, and ReelBench demonstrate improvements in content coverage and visual consistency over existing baselines.

Critical strengths
  • The paper writing is clear, and the motivation of improving long-horizon video generation through evolving agentic workflows is reasonable.
  • The proposed experience memory and multimodal visual memory provide a clean framework for improving script generation and maintaining cross-shot consistency.
  • The experimental evaluation is comprehensive. The authors also introduce ReelBench as a benchmark.
Critical weaknesses
  • The contribution of the proposed visual memory is not sufficiently validated. The paper lacks ablations comparing different memory configurations, such as static references, previous-frame conditioning, and the proposed retrieval strategy.
  • The inference efficiency analysis is insufficient. The framework involves multiple agents and iterative video generation, but the paper lacks comparisons of inference latency and the number of model calls against existing methods.
  • The related work section could be refined to further position the paper with respect to recent video generation methods on selective memory and consistency designs, such as
    • OneStory: Coherent multi-shot video generation with adaptive memory (CVPR 2026)
    • VGGRPO: Towards World-Consistent Video Generation with 4D Latent Reward (ECCV 2026)
Questions for authors

Please see the weakness above.

Comments on the paper

1 comment

Hi Reviewers,

The authors have posted their rebuttal along with new experimental results. Please review their responses, engage with the authors to discuss any remaining concerns. Best,

Area Chair

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.