PaperlayerView exact revision on arXiv

Source-linked AI summary

Presenting a Paper is an Art: Self-Improvement Aesthetic Agents for Academic Presentations

Chengzhi Liu, Yuzhe Yang, Kaiwen Zhou, Zhen Zhang, Yue Fan, Yanan Xie, Peng Qi, Xin Eric Wang

arXiv:2510.05571v2cs.CL

TL;DR

Existing automated presentation methods have limitations in storytelling, aesthetic quality, and self-adjustment. EvoPresent combines agent-based generation with PresAesth feedback and a benchmark for presentation quality and aesthetic awareness. The findings highlight stronger aesthetic-task generalization with multi-task RL and the importance of feedback quality for refinement.

  • Problem

    Existing automated methods have limitations in narrative coherence, design flexibility, self-improvement, and reliable aesthetic evaluation.

    Source: §1 · p. 1

  • Method

    EvoPresent uses a draft–feedback–refinement agent framework with PresAesth, a multi-task RL model for aesthetic scoring, defect correction, and comparison.

    Source: §1 · pp. 1–2

  • Results

    Multi-task RL generalizes better across aesthetic-awareness tasks; PresAesth’s scoring MAE is about 18% lower than that of GPT-4o and Claude-4-sonnet.

    Source: §1 · p. 2, §5.1 · p. 7

  • Takeaways & Limitations

    High-quality feedback supports faster iterative refinement, while initial performance alone does not guarantee effective self-correction.

    Source: §1 · p. 2

Abstract

from arXiv · show

The promotion of academic papers has become an important means of enhancing research visibility. However, existing automated methods struggle limited storytelling, insufficient aesthetic quality, and constrained self-adjustment, making it difficult to achieve efficient and engaging dissemination. At the heart of those challenges is a simple principle: \emph{there is no way to improve it when you cannot evaluate it right}. To address this, we introduce \textbf{EvoPresent}, a self-improvement agent framework that unifies coherent narratives, aesthetic-aware designs, and realistic presentation delivery via virtual characters. Central to EvoPresent is \textbf{PresAesth}, a multi-task reinforcement learning (RL) aesthetic model that provides reliable aesthetic scoring, defect adjustment, and comparative feedback, enabling iterative self-improvement even under limited aesthetic training data. To systematically evaluate the methods, we introduce \textbf{EvoPresent Benchmark}, a comprehensive benchmark comprising: \textit{Presentation Generation Quality}, built on 650 top-tier AI conference papers with multimodal resources (slides, videos and scripts) to assess both content and design; and \textit{Aesthetic Awareness}, consisting of 2,000 slide pairs with varying aesthetic levels, supporting joint training and evaluation on scoring, defect adjustment, and comparison. Our findings highlight that (i) High-quality feedback is essential for agent self-improvement, while initial capability alone does not guarantee effective self-correction. (ii) Automated generation pipelines exhibit a trade-off between visual design and content construction. (iii) Multi-task RL training shows stronger generalization in aesthetic awareness tasks.

1 INTRODUCTION

Existing automated presentation methods have gaps in narrative coherence, design flexibility, self-improvement, and reliable aesthetic evaluation. EvoPresent combines iterative generation with aesthetic feedback, and its evaluation highlights the importance of feedback quality and multi-task learning.

1 INTRODUCTION

  • Existing approaches are limited by weak narrative coherence, fixed-template design, absent self-improvement, and unreliable aesthetic evaluation.Evaluation also relies heavily on VLM-as-judge methods, which the authors say reduces consistency and robustness.Source: §1 · p. 1
  • EvoPresent uses a draft–feedback–refinement loop to create presentations with coherent narratives and aesthetic-aware design.Its Storyline, Scholar, Design, and Checker Agents construct scripts, enrich content, render presentations, and provide targeted feedback.Source: §1 · pp. 1–2
  • EvoPresent Benchmark combines 650 annotated academic resources in slides, videos, and scripts with 2,000 perturbed slide pairs for aesthetic-task training and evaluation.Generation quality is assessed globally and across eight localized content and design dimensions; the slide pairs support scoring, defect adjustment, and comparison.Source: §1 · p. 2
  • Multi-task RL generalizes better across aesthetic-awareness tasks, while high-quality feedback supports faster refinement and initial performance alone does not ensure effective correction.Automated generation also reveals a trade-off between content construction and visual design, with aesthetic awareness identified as a bottleneck.Source: §1 · p. 2
  • The authors describe EvoPresent as a multi-agent framework for realistic presentations with minimal human intervention, alongside PresAesth and a benchmark for systematic evaluation.The stated contributions include higher-quality, more engaging presentations comparable to human-designed ones.Source: §1 · p. 2

2 RELATED WORK

Prior work on automated presentation generation and aesthetic evaluation leaves gaps in narrative coherence, design flexibility, self-refinement, and academic-slide aesthetic awareness. EvoPresent combines storyline construction and iterative refinement with PresAesth for academic visual design.

2 RELATED WORK

  • EvoPresent’s framework combines storyline construction, iterative self-refinement, and an aesthetic checker trained on human preferences for scoring, defect adjustment, and comparison.The framework overview depicts the checker guiding iterative improvement after presentation design and rendering.Source: Figure 2 · p. 3
  • PPTAgent and PresentAgent rely on direct extraction and fixed templates, while Paper2Poster’s VLM checker has limited aesthetic perception; these approaches lack reliable self-improvement.The passage contrasts EvoPresent’s storyline construction and iterative self-refinement with these limitations (Li et al., 2023).Source: §2 · p. 3
  • Existing aesthetic-evaluation methods largely target natural images and struggle with the subjectivity and complexity of academic slides (Zhou et al., 2024b).Prior work explores image-design scoring (Hong et al., 2023; Yu et al., 2021) and preference alignment.Source: §2 · p. 3

3 EVOPRESENT FRAMEWORK

EvoPresent generates academic presentations through a draft–feedback–revision cycle, with agents handling storyline construction, content enrichment, design, and checking. PresAesth supplies aesthetic judgments and actionable feedback for iterative refinement.

3 EVOPRESENT FRAMEWORK

  • EvoPresent sequences four agents in a draft–feedback–revision cycle, with PresAesth supporting aesthetic scoring, defect correction, and pairwise comparison.The Checker Agent uses the model’s feedback to guide iterative improvement.Source: §3 · p. 3

3.1 EVOPRESENT WORKFLOW

  • The Storyline Agent organizes paper text and visuals into thematic sections, coherent chapters, and a complete presentation script.It extracts key arguments and details, associates relevant visuals and formulas, and filters irrelevant elements.Source: §3.1 · p. 4
  • The Scholar Agent enriches the storyline with related knowledge and generated charts or diagrams to improve visual impact and align visuals with text.It uses knowledge-access tools and image-generation models.Source: §3.1 · p. 4
  • The Design Agent plans balanced layouts, renders a selected visual style, adjusts font hierarchy, and adds dynamic effects; HTML is chosen for flexibility and aesthetic control.Its Layout Planner uses text and image sizes and aspect ratios to establish a stable initial layout before rendering.Source: §3.1 · p. 4
  • The Checker Agent stops when a slide meets the aesthetic threshold, otherwise feeds back revisions and retains the highest-scoring version if quality degrades or the target is unmet.Final slides are paired with narration audio aligned to each slide; video transitions are optional.Source: §3.1 · p. 4

3.2 PRESAESTH: MULTI-TASK AESTHETIC AWARENESS MODEL

  • PresAesth scores slides, identifies defects with improvement feedback, and compares revisions; it is trained using GRPO with verified rewards for format and task accuracy.The reward formulation parses feedback categories and uses tolerance thresholds for F1-score and scoring error.Source: §3.2 · pp. 4–5

4 EVOPRESENT BENCHMARK

The benchmark evaluates presentation generation on content and design and evaluates aesthetic awareness through scoring, defect adjustment, and comparison. Its two components pair annotated multimodal paper presentations with perturbed slide pairs.

4.1 TASK DEFINITION

  • Generation-quality evaluation measures global content and design and uses VLM judgments across eight localized dimensions; aesthetic-awareness evaluation uses MAE, F1-score, and accuracy for its three tasks.Localized dimensions cover fidelity, clarity, narrative, engagement, elements, layout, hierarchy, and color.Source: §4.1 · p. 5

4.2 BENCHMARK CONSTRUCTION

  • The generation-quality collection covers 650 top-conference papers across domains, each with slides, videos, and scripts annotated by two to three experts.The aesthetic-awareness suite perturbs original slides into poor, base, and good variants, then forms pairs with independent quality ratings and defect labels.Source: §4.2 · pp. 5–6
  • The aesthetic-awareness dataset contains 2,000 slide pairs, split into 1,600 training and 400 testing pairs.Presentations average 9.5 minutes, range from 6–19 slides, and include slides with at least 2–3 design errors across categories.Source: §4.2 · p. 6, Figure 3 · p. 6

5 EXPERIMENTS

EvoPresent performs strongly across presentation quality, aesthetic awareness, and delivery evaluations, while ablations highlight the roles of agent feedback and multi-task training in refinement.

5 EXPERIMENTS

  • Experiments compare oracle, end-to-end, multi-agent, and EvoPresent generation methods, alongside closed- and open-source models for aesthetic-awareness evaluation.The benchmark comparison covers presentation generation and evaluation.Source: Table 1 · p. 6, §5 · p. 6

5.1 EVALUATION AND ANALYSIS

  • About 17% lower perplexity for EvoPresent-4o accompanies stronger content and visual-design results, though visual-design gains can trade off against content construction.Reasoning models can produce better visual quality but introduce redundancies and higher perplexity; most models differ more in aesthetics and design than in content quality.Source: §5.1 · pp. 6–7, Table 2 · p. 7
  • EvoPresent receives the best video-performance scores across four dimensions and achieves higher accuracy on both verbatim and explanatory content questions.Video ratings use a 1–5 scale; content questions are answered by Qwen2.5-VL-32B and GPT-4o on 30 sampled papers.Source: Figure 4 · p. 7, §5.1 · p. 8
  • PresAesth performs best across three aesthetic-awareness tasks, with scoring MAE about 18% lower than closed-source models and accuracy of 87.8%, about 40% higher than other models.Its structured reasoning also makes feedback more explainable.Source: Table 3 · p. 7, §5.1 · pp. 7–8
  • EvoPresent’s slides show clearer hierarchy and multi-page narrative flow, and human evaluators rate its generation quality above other methods and competitive with human presentations.The comparison highlights overlap, content, typography, and layout defects in generated variants.Source: Figure 5 · p. 8, §5.1 · pp. 8–9, Figure 6 · p. 8

5.2 ABLATION STUDY

  • Removing the Scholar Agent lowers content score from 3.91 to 3.40; removing the Design Agent cuts design by 10.2%, while removing the Checker cuts design by 5.1% and aesthetics by 15%.The ablation attributes content coverage and narrative depth to Scholar, layout balance and readability to Design, and detail alignment and visual consistency to Checker.Source: Table 4 · p. 9, §5.2 · p. 9
  • Single-task GRPO performs worse than PresAesth, while multi-task SFT exceeds single-task RL accuracy but returns static labels rather than actionable feedback.The authors suggest aesthetic tasks share principles such as balance and visual consistency; GRPO generates targeted feedback aligned with agent self-improvement.Source: Table 5 · p. 9, §5.2 · p. 9
  • With PresAesth as checker, EvoPresent-4o’s aesthetic score rises from 3.2 to over 8.0 within three iterations; other checkers need more than five iterations and finish lower.Higher initial accuracy alone does not ensure stronger self-correction: GPT-4o and DeepSeek-R1 gain under 20% in 1–2 iterations, and stronger initial models also show limited correction.Source: Figure 7 · p. 9, §5.2 · p. 9

6 CONCLUSION

EvoPresent combines iterative self-improvement with a benchmark for presentation generation and aesthetic awareness, and uses multi-task GRPO to train PresAesth.

6 CONCLUSION

  • EvoPresent uses iterative aesthetic-aware optimization to generate presentations aligned with human preferences.The framework is built on a benchmark for evaluating presentation-generation quality and integrated training and evaluation of aesthetic-awareness tasks.Source: §6 · p. 9

A.1 METRICS

The benchmark evaluates presentation content and design through narrative-coherence, content-fidelity, and visual-design measures grounded in perceptual principles and visual hierarchy.

A.1 METRICS

  • The selected benchmark indicators are grounded primarily in Gestalt perceptual principles and visual hierarchy theory to quantify slide-design flaws.The section also references visual-design and perception theory literature.Source: §A.1 · p. 14
  • Perplexity evaluates narrative coherence from each slide and its predecessor; lower scores indicate better narrative flow and contextual coherence.The evaluation uses Qwen2.5-VL-7B.Source: §A.1 · p. 14
  • ROUGE-L measures the longest common subsequence between generated scripts and reference content, with higher scores indicating better content preservation.The metric uses balanced precision and recall when β is typically set to 1.Source: §A.1 · p. 14
  • Layout Balance assesses visual equilibrium by segmenting slide elements and calculating their area-weighted center of mass.Its score reflects the distance between that center and the slide center.Source: §A.1 · p. 14
  • The benchmark also assesses fidelity, clarity, narrative, and engagement across content accuracy, comprehensibility, story flow, and audience attention.These criteria evaluate whether slides represent source information faithfully, communicate clearly, form a coherent story, and hold attention.Source: §A.1 · pp. 14–15
  • Visual-design measures cover the quality and appropriateness of elements, spatial arrangement, information hierarchy, and color use.The measures consider element quality and relevance, balanced use of space, distinction between primary and secondary information, and purposeful, readable color.Source: §A.1 · p. 15

A.2 DATA CONSTRUCTION

Aesthetic-awareness data pairs real slides with human-annotated variants that differ in design quality through predefined modifications.

A.2 DATA CONSTRUCTION

  • Variants are generated through within-object alignment changes, between-object layout changes, and typography alterations.These modifications change internal element structure, spatial relationships between elements, or font size, weight, and spacing.Source: §A.2 · p. 15
  • The dataset combines real slide data with augmented variant images that may improve or degrade design quality.Human annotators assess and annotate the resulting variants to establish the evaluation benchmark.Source: §A.2 · p. 15
  • The resulting slide pairs provide test samples for evaluating design improvement and defect identification.The variants include both positive changes that may enhance visual effects and negative changes that may reduce design quality.Source: §A.2 · p. 15

B MULTI-TASK AESTHETIC AWARENESS RL TRAINING

The GRPO training procedure compares sampled responses within each query group, normalizes their quality differences, and optimizes the policy with a stability-regularizing KL term.

B MULTI-TASK AESTHETIC AWARENESS RL TRAINING

  • For each query, GRPO samples N = 8 responses from the old policy and obtains a reward for each from the reward model.Source: §B.1 · p. 15
  • GRPO normalizes each response's quality relative to the other responses in its group, so learning uses comparative rather than absolute quality.Source: §B.1 · p. 15
  • The optimization objective uses a policy ratio for response probabilities and a KL-divergence term that regularizes training stability by limiting deviation from a reference model.Source: §B.1 · p. 16

B.2 TRAINING SETTINGS

The appendices describe PresAesth training settings and examples, the framework’s presentation visualizations, and a modular pipeline for producing narrated presentation videos.

B.2 TRAINING SETTINGS

  • AesthEval starts from Qwen-VL-7B and trains for two epochs on eight NVIDIA H100 GPUs, sampling eight responses per query during GRPO rollouts.Hyperparameters were selected using a held-out validation set to provide meaningful, sufficiently dense reward signals.Source: §B.2 · p. 16
  • Training uses approximately 3,400 human-annotated instances: 500 for comparison, 1,900 for defect detection, and 1,000 for scoring.The data combines academic presentations from EvoPresent Benchmark with general-domain slides from SlideAudit.Source: §B.2 · p. 16
  • Single-slide adjustment and scoring inputs are resized to 960 pixels on the longest edge, while three-image comparison inputs use 720 pixels.The aspect ratio is maintained; the smaller comparison resolution manages its increased computational load.Source: §B.2 · p. 16

C.1 PRESAESTH EXAMPLE OUTPUTS

  • The example outputs show AesthEval scoring a slide and comparing two slides, with one example describing Slide B as cleaner and more engaging.A separate example gives a slide an overall quality score of 8.50.Source: §C.1 · p. 17

C.2 PRESENTATION GENERATION

D.1 PRESENTATION VIDEO GENERATION

  • The video pipeline synthesizes narration, animates a reference face, composites it with rendered slides, and fuses the audio and visuals into a final video.It supports either MegaTTS3 or OpenAI TTS, and can produce narration-only, background-only, or complete face-driven presentations.Source: §D.1 · p. 24

E.1 CHECKLIST

The appendices outline human annotation procedures and checklists, then specify prompts and scoring rubrics for assessing presentation quality.

E.1 CHECKLIST

  • Approximately 30 design-background volunteers annotate each slide independently, with two to three annotators per slide to reduce subjective bias.The process uses a Gradio interface and combines scoring, deficiency identification, and consistent evaluation criteria.Source: §E.1 · p. 24
  • The annotation checklist covers content clarity and organization alongside layout, color, typography, image quality, readability, and audience engagement.It is intended to capture quantitative scores and qualitative feedback for training and evaluating the slide design assessment model.Source: §E.1 · pp. 24–26

F.1 TRAINING PROMPTS

  • Training prompts request a 1-to-10 slide rating, category-specific deficiency diagnosis and improvements, or a choice between two enhanced slides.The comparison prompt considers fidelity and consistency, perceptual quality, clarity and readability, and communication effectiveness.Source: §F.1 · pp. 26–27

F.2 GENERATION QUALITY EVALUATION PROMPTS

  • Generation-quality prompts assess fidelity, clarity, narrative, engagement, visual elements, layout, and hierarchy using dimension-specific criteria and five-point rubrics.The fidelity rubric directs evaluators to use a reference when available and to ignore style unless it obscures truthfulness.Source: §F.2 · pp. 27–28
  • The narrative prompt evaluates story arc, cohesion, progression, and framing across ordered slides and an optional transcript.Its rubric distinguishes a strong arc with seamless transitions and explicit takeaways from fragmented or disjoint presentations.Source: §F.2 · p. 28
  • The visual prompts judge asset relevance and quality, grid alignment and spacing, and whether visual cues establish hierarchy and reading order.The asset, layout, and hierarchy rubrics each describe performance levels from 1 to 5.Source: §F.2 · pp. 28–29

Related papers

Loading 2510.05571v2…