acceptodds
Withdrawn from ICLR 2027

OmniPresent: Generating Coherent Presentation Suites with Self-Improving Planners

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Transforming static research papers into dynamic media such as posters, slides, and videos is essential for effective dissemination but remains a labor-intensive challenge. Existing automated approaches often treat these formats in isolation and consequently fail to maintain semantic consistency across the entire presentation suite. We address this fragmentation by formalizing the task of unified presentation suite generation and proposing to orchestrate the creation of coherent deliverables. Our framework adopts a renderable HTML representation to enable centralized content planning and a self-correcting verify-and-repair loop that actively resolves conflicts across modalities. It further learns from verification experience across papers, converting recurring, evidence-grounded failures into reusable updates to format-specific planners for future papers. We further facilitate scalable research in this domain by releasing , a comprehensive dataset comprising over one thousand papers with paired artifacts, and establishing a rigorous VLM-based evaluation protocol. Empirical results confirm that our method generates high-quality and faithful presentation suites that greatly surpass strong baselines in both accuracy and visual appeal. Code and data will be released.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Multi-agent System, Unified Presentation Generation, Academic Media, Cowork Agent

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_Rr5gOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes OmniPresent, a unified framework for generating coherent presentation suites, including posters, slides, videos, and project pages, from a single research paper.

It introduces a shared Knowledge Stream and Cross-Artifact Verification to detect and correct factual inconsistencies and missing content across formats.

It proposes Experiential Planner Learning (EPL), which leverages accumulated verification failures to improve planner instructions for future papers.

It introduces OmniPreBench, a benchmark comprising 1,036 papers with paired presentation artifacts, along with an LLM/VLM-as-a-Judge evaluation protocol.

Critical strengths

The paper addresses the practical and important problem of cross-format consistency, moving beyond existing artifact-specific generation approaches.

The proposed self-improving framework leverages past verification experience to refine planners, rather than merely correcting individual outputs.

The paper provides extensive comparisons across multiple backbone models and baselines, supported by CV/EPL ablation studies and human evaluation.

Critical weaknesses

Key components of OmniPresent have already been explored in prior work, including unified multi-format generation in PaperX, feedback-driven self-improvement in EvoPresent, and iterative multi-agent refinement in Paper2Poster. The proposed framework primarily combines these existing ideas, with limited methodological novelty beyond their integration.

EPL primarily updates planner instructions (prompts), but concrete examples of these updates and analyses explaining their effectiveness are lacking. Moreover, improvements in final FactScore and QA are not consistently statistically significant.

(Limitations of Cross-Artifact Verification) The framework uses Slides as the backbone for checking content coverage across Poster and Page, while direct verification of Video and post-rendering outputs is limited. The assumption that Slides should cover all salient content from other formats may also be overly restrictive.

The contributions of individual components, particularly the Shared Knowledge Stream, are not adequately isolated, making it difficult to determine their respective effectiveness.

Questions for authors

I would appreciate it if the authors could address the concerns raised in the weaknesses above.

Reviewer_k53pOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper studies generating a consistent suite of slides, a poster, a project webpage, and a narrated video from a research paper. OmniPresent combines shared evidence-grounded content, format-specific planners, cross-artifact verification and repair, and HTML rendering. Experiential Planner Learning (EPL) further updates planner instructions using failures from previous papers. The authors introduce OmniPreBench, containing 1,036 papers with paired artifacts, and evaluate on 50 papers using automated metrics and a 13-participant visual-quality study. The results show improved information coverage and visual quality, along with fewer repair rounds after EPL.

Critical strengths

S1. The paper addresses a useful problem with a coherent system and benchmark. Sharing source evidence across presentation formats and checking their consistency is a sensible response to the errors introduced by separate generation pipelines. The paired benchmark provides a useful basis for studying this task. Improvements across several backbones and the human visual-quality results support the system’s practical value.

S2. EPL has meaningful evidence of transfer to held-out papers. The authors separate Meta-Train, Meta-Dev, and Final-Test, freeze the selected planners before testing, and report paired uncertainty estimates. On Final-Test, repair rounds decrease from 2.42 to 1.66 for Gemini and from 2.30 to 1.58 for GPT-5-mini, with confidence intervals excluding zero. This supports the claim that planner updates can reduce subsequent repair effort.

Critical weaknesses

W1: The main factual-faithfulness claim is not adequately validated. Equation (7) measures the fraction of source-paper facts covered by an artifact. It does not directly penalize additional unsupported claims, so a higher score need not mean fewer hallucinations. QA also primarily measures information preservation. An independent assessment of unsupported claims and cross-format contradictions is needed to establish that the system improves factual reliability.

W2: The contribution of cross-format coordination is not sufficiently isolated. Removing CV removes source-evidence checking, gap filling, and iterative repair together with cross-format checking. This establishes the benefit of the combined module, but leaves unclear how much requires joint generation rather than comparable verification of each artifact independently.

Questions for authors

Q1: Can you independently assess unsupported claims and cross-format contradictions in the final rendered artifacts, using human annotations or a separately validated evaluator? How well does the pipeline’s Conflict Check agree with this assessment?

Q2: How does joint CV compare with independent per-artifact verification using the same source evidence, backbone, repair budget, and output constraints? Please specify the actual repair limit and the generation costs. This comparison would directly affect my assessment of the cross-format contribution.

Reviewer_ijw2OpenReview ↗
Rating
1
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes OmniPresent, a unified framework for generating coherent scientific presentation suites of slides, posters, videos, and webpages. OmniPresent employs iterative verification and repair loops to resolve inconsistencies across generated artifacts. The paper also introduces OmniPreBench, a dataset comprising over 1,000 papers paired with presentation artifacts. Experimental results demonstrate that OmniPresent generates high-quality presentation suites.

Critical strengths
  • The paper addresses a practically meaningful problem of unified scientific presentation generation.
  • The curation of a benchmark comprising research papers paired with presentation artifacts across multiple formats is a valuable contribution.
  • The proposed method demonstrates strong performance across the reported LLM-based evaluation metrics.
Critical weaknesses
  • Although prior works [1, 2] are already cited, the novelty of OmniPresent over these existing unified presentation generation frameworks remains unclear. A more detailed discussion of the technical differences and contributions would strengthen the paper.

  • Cross-artifact conflict and coverage gap metrics are reported only for the proposed method. Comparing these metrics against existing baselines would provide stronger evidence of the framework's effectiveness in resolving cross-artifact inconsistencies.

  • The main evaluation includes only one or two specialized baselines for each artifact type. Comparisons with more recent methods would strengthen the experimental evaluation.

Questions for authors
  • Could the authors clarify the key technical differences and novel contributions of OmniPresent compared with prior works [1, 2]?

  • Could the authors provide comparisons of cross-artifact conflicts and coverage gaps against existing baselines, particularly [2] and human-authored artifacts?

  • Several more recent methods [1, 3, 4] are neither discussed in the related work nor included in the experimental comparisons. Could the authors include additional baselines in the experimental comparisons?

  • The human preference study guidelines appear to cover both content quality and visual quality, but only visual-quality results are reported. Could the authors clarify whether content-quality ratings were collected and, if so, provide the corresponding results?

[1] Lingao Xiao, et al. ResearchStudio-Reel: Automate the Last Mile of Research from Paper to Poster, Video, and Blog. arXiv preprint arXiv:2607.04438, 2026. URL: https://arxiv.org/abs/2607.04438.

[2] Tao Yu, et al. PaperX: A Unified Framework for Multimodal Academic Presentation Generation with Scholar DAG. 2026. URL: https://arxiv.org/abs/2602.03866.

[3] Tarik Can Ozden, et al. Narrative-Driven Paper-to-Slide Generation via ArcDeck. In Proceedings of the European Conference on Computer Vision (ECCV), pages 520–537, 2026.

[4] Hao Zheng, et al. DeepPresenter: Environment-Grounded Reflection for Agentic Presentation Generation. In Findings of the Association for Computational Linguistics (ACL), pages 31545–31558, 2026. URL: https://aclanthology.org/2026.findings-acl.1578/.

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

Related papers

Discussion (0)

Sign in to comment.