PaperlayerView exact revision on arXiv

Source-linked AI summary

Building Presentation Intelligence for High-Quality Slide Generation

Yuhang Xie, Cheng Li, Chenliang Li, Mingyu Xu, Mingwei Sun, Suhang Zheng, Ming Yan, Xiaoxiao Xu, Lin Qu

arXiv:2610.30888v1cs.LG

TL;DR

High-quality PPT generation requires jointly addressing content and visual aesthetics, while data construction, supervision, and reward design are difficult across domains. The paper combines cross-validated synthesis, teacher enhancement, and verifiable-reward reinforcement learning; its SFT models improve benchmark performance and RL improves aesthetic metrics. The results support the effectiveness of this training framework for PPT generation.

  • Problem

    PPT generation must jointly optimize content and visual aesthetics, which are subjective and domain-sensitive and complicate data construction, supervision, and reward design.

    Source: §1 · pp. 1–2

  • Method

    The framework combines cross-validated data synthesis, guided teacher enhancement, mixed-teacher SFT, outcome-driven trajectory filtering, and GRPO with verifiable aesthetic rewards.

    Source: §1 · pp. 2–3, §7 · p. 10

  • Results

    The pipeline improves content and aesthetics; on PresentBench, Qwen-3.6-27B and Qwen-3.6-35B-A3B gain 7.45 and 9.35 points, respectively, while RL improves five aesthetic metrics.

    Source: §1 · p. 3, §5.3 · p. 7, §6.5 · p. 9

  • Takeaways & Limitations

    The results demonstrate the effectiveness of a unified training framework for high-quality PPT generation.

    Source: §1 · p. 3, §7 · p. 10

    Overly aggressive trajectory filtering can remove difficult informative samples, reduce data diversity, and occasionally degrade performance.

    Source: §5.3 · p. 7

Abstract

from arXiv · show

PowerPoint presentations (PPTs) are a dominant medium for communication across diverse domains, yet high-quality PPT generation still depends heavily on labor-intensive human effort. Recent advances in large language models (LLMs) for agentic tasks have created both new opportunities and challenges (Chen et al., 2026) for automatic PPT generation. In this work, we present a unified training framework for automatic PPT generation toward presentation intelligence, integrating cross-validated data synthesis, guided teacher enhancement, and verifiable reward optimization. We first construct a cross-validated pipeline for data synthesis, with explicit control over both difficulty and validity. We then develop a teacher-enhancement framework grounded in rules and prior knowledge to substantially strengthen the instructional capacity of teacher models. Building on this, we further improve supervised fine-tuning (SFT) through a mixed-teacher strategy and a lightweight outcome-driven trajectory filtering method. Finally, based on the resulting SFT models, we introduce a suite of verifiable aesthetic metrics together with a GRPO-based reinforcement learning framework to optimize slide-level visual attributes beyond text-only supervision. Experiments show that our enhanced SFT pipeline improves Qwen-3.6-27B and Qwen-3.6-35B-A3B by 7.45 and 9.35 points on PresentBench, respectively, surpassing GLM-5.2 while maintaining strong generalization. Reinforcement learning further improves performance across five aesthetic dimensions, compensating for the limitations of SFT in visual refinement. Overall, the results demonstrate the effectiveness of our framework for high-quality PPT generation.

1 Introduction

The introduction frames PPT generation as a joint content-and-aesthetics problem and presents a training pipeline spanning data synthesis, teacher enhancement, and verifiable rewards. Experiments report gains on PresentBench and improvements in slide aesthetics.

1 Introduction

  • PPT generation must jointly optimize content quality and visual aesthetics, which are subjective and domain-sensitive.Academic slides prioritize rigor and precision, while advertising slides favor attractiveness and persuasive impact.Source: §1 · pp. 1–2
  • Prior approaches give comparatively limited attention to principled data synthesis, task-specific training, and comprehensive verifiable rewards.The introduction notes that existing work often lacks systematic data-construction accounts and explicit assessment of training-data difficulty and quality.Source: §1 · p. 2
  • The framework combines cross-validated data synthesis, rule- and knowledge-grounded teacher enhancement, mixed-teacher SFT, trajectory filtering, and GRPO with verifiable aesthetic rewards.The rewards target slide-level visual quality beyond text-only supervision.Source: §1 · pp. 2–3
  • Experiments report that, in the SFT stage, the method improves Qwen-3.6-27B and Qwen-3.6-35B-A3B by 7.45 and 9.35 points, respectively.With only 4,000+ training samples, Qwen-3.6-27B surpasses the substantially larger GLM-5.2; in the RL stage, both models gain across five aesthetic metrics.Source: §1 · p. 3

2 Related Work

Prior work advances agentic slide-generation frameworks, but training strategies for end-to-end PPT generation remain underdeveloped. Existing approaches vary in planning, supervision, and reward verification.

2.1 LLM-Based Slide Generation Frameworks

  • Agentic systems use multi-stage planning, iterative editing, collaborative multimodal drafting, or environment-grounded reflection, but still rely largely on foundation-model capabilities.Examples include PPTAgent, PreGenie, and DeepPresenter.Source: §2.1 · p. 3

2.2 Training Strategies for Slide Generation

  • Training approaches use SFT or reinforcement learning, yet rely on heuristic data construction, lack task-specific teacher enhancement, or use unverified rewards.AeSlides focuses on next-slide generation rather than full-deck generation, leaving end-to-end training methods relatively underexplored.Source: §2.2 · p. 3

3 Presentation Generation Method Selection

The paper compares code-centric presentation-generation paradigms and evaluates them across generation settings. It selects Single-HTML, which performs best overall in the reported evaluation.

3 Presentation Generation Method Selection

  • Image-based generation often has textual hallucinations, weaker controllability, and greater computational cost, so the study focuses on code-centric paradigms.The compared paradigms are Python-based code generation, HTML rendering, and direct image generation.Source: §3 · p. 4
  • The evaluation covers two Qwen backbones, PresentBench and GDPVal, and single-page and multi-page generation, using a unified per-rubric evaluation protocol.HTML outputs are rendered through browser screenshots, while Python outputs are converted to PDF with headless LibreOffice.Source: §3 · p. 4
  • Single-HTML consistently delivers the best overall performance for both Qwen-3.6-27B and Qwen-3.6-35B-A3B, leading the authors to adopt it for subsequent training.HTML outputs outperform Python outputs in visual design, including layout, color, diversity, and mathematical rendering.Source: §3 · p. 4

4 Cross-Validated Data Synthesis

The data-synthesis pipeline uses cross-stage constraints to control task difficulty and validity. It links source evidence, tasks, and rubrics to avoid self-certification and support auditing.

4 Cross-Validated Data Synthesis

  • The pipeline constructs tasks and instance-level rubrics under cross-stage constraints, making task difficulty, rubric coverage, and validity explicitly controllable.It produced 4,802 instances with 209,966 rubric items across 25 primary and 162 secondary domains.Source: §4 · p. 4
  • Source cards ground synthesis; tasks withhold answer-critical evidence, rubrics separately verify completeness and correctness, and coupled audits enforce consistency and evidential coverage.The final rubric is tightened to increase discriminative power without changing its measurement structure.Source: §4 · p. 5

5 Guided Supervised Fine-Tuning

The SFT pipeline strengthens distilled supervision through iterative and strong-to-weak guidance, teacher mixing, and outcome-driven filtering. It improves performance across two model backbones and benchmarks, while the case study identifies visual defects that remain after SFT.

5.1 Data Distillation Method

  • Self-guided distillation iteratively analyzes teacher failures and induces principles for redistillation, while strong-to-weak guidance transfers stronger models’ planning, grounding, and organization strengths to weaker teachers.The approach aims to improve weaker, more cost-efficient teachers’ supervision and reduce overall training cost.Source: §5.1 · p. 5

5.3 SFT Results

  • 77.90 (+7.45) on PresentBench for Qwen-3.6-27B surpasses GLM-5.2 (75.04), while both fine-tuned Qwen backbones also improve on GDPVal.Qwen-3.6-35B-A3B reaches 74.28 (+9.35) on PresentBench; the best fine-tuned Qwen-3.6-27B and Qwen-3.6-35B-A3B reach 79.36 (+4.87) and 78.17 (+4.16) on GDPVal, respectively, both exceeding GLM-5.2 (74.93).Source: Table 2 · p. 6, §5.3 · p. 7

5.2 Training Method

  • The SFT corpus combines teacher source and composition strategy, with Mixed+Filter adding rule-based removal of rendering failures, overflow, overlap, and related issues.Single, Mixed, and Mixed+Filter regimes respectively isolate a supervisory source, combine teacher corpora, and filter the mixed corpora.Source: §5.2 · p. 6

5.3 SFT Results

  • Guided variants improve scores over the basic prompt from 70.52 to 71.97 for Qwen-3.6-27B and from 67.29 to 68.73 for Qwen-3.6-35B-A3B.The passage attributes the gains to stronger structural regularity and richer intermediate supervision in teacher trajectories.Source: §5.3 · p. 7
  • Teacher mixing raises PresentBench scores from 76.93 to 77.90 (+0.97) for Qwen-3.6-27B and from 69.55 to 73.49 (+3.94) for Qwen-3.6-35B-A3B.The authors attribute the gain to combining Qwen-3.8-Max’s broad reasoning and planning with GLM-5.2’s presentation-specific priors.Source: §5.3 · p. 7
  • Filtering raises the mixed-teacher Qwen-3.6-35B-A3B PresentBench score from 73.49 to 74.28 (+0.79), but overly aggressive filtering can remove difficult informative samples and occasionally degrade performance.The passage frames trajectory curation as a trade-off between high-quality supervision and coverage of challenging patterns.Source: §5.3 · p. 7

5.4 SFT Case Study

  • SFT improves content organization and visual presentation, but excessive whitespace, visual imbalance, and element overlap remain and motivate visual-reward reinforcement learning.The described examples show clearer hierarchy, richer visual structures, and better text–graphic coordination after SFT.Source: §5.4 · p. 7

6 Verifiable RL Training

The RL stage initializes from selected SFT checkpoints, then uses verifiable aesthetic rewards and GRPO to refine slide visuals. Results improve across five aesthetic dimensions, with particularly strong gains for the MoE model.

6.1 Initialization Model Selection

  • The selected post-SFT checkpoints preserve GDPVal PPT-subset performance while attaining the highest PresentBench score among candidate models.Qwen-3.6-27B and Qwen-3.6-35B-A3B initialize the dense and MoE policies, respectively.Source: §6.1 · p. 7

6.2 Reward Design

  • Five deterministic, interpretable metrics score rendered slides for aspect ratio, whitespace, collisions, balance, and font readability.Their normalized scores are combined by an unweighted mean into slide-level and deck-level rewards.Source: §6.2 · pp. 7–8

6.3 RL Data Selection

  • RL training uses task-hard queries from low-scoring benchmark-rubric outputs and backbone-specific aesthetic-hard examples with remaining visual deficiencies.Together, these datasets cover globally difficult tasks and model-specific aesthetic failure modes.Source: §6.3 · p. 8

6.4 RL with Verifiable Rewards

  • The enhanced GRPO framework combines reward-decoupled normalization, token-level optimization, and stabilized clipping; only renderable presentations receive aesthetic rewards.Training also studies both use and omission of KL regularization; each update uses 64 rollouts and training lasts one epoch.Source: §6.4 · p. 8

6.5 RL Result

  • 4.36 on PresentBench and 3.75 on the GDPVal PPT subset are the best Qwen-3.6-35B-A3B overall scores, outperforming no-train and SFT baselines.RL improves aesthetic quality across all five metrics; Figure 3 illustrates three cases.Source: §6.5 · pp. 8–9, Table 3 · p. 9, Figure 3 · p. 9
  • Task-hard data consistently outperforms aesthetic-hard data on both backbones, which the authors attribute to broader coverage of content and layout challenges.Aesthetic-hard targets narrower failure modes and is more susceptible to weak advantage signals and local overfitting.Source: §6.5 · p. 9
  • Removing KL regularization is consistently equal to or better than retaining it on PresentBench under the same data-selection strategy.With verifiable rewards and render-validity constraints, KL can restrict movement toward higher-reward layouts.Source: §6.5 · pp. 9–10

7 Conclusion

The conclusion presents a unified training framework combining cross-validated data synthesis, teacher enhancement, and verifiable-reward RL for PPT generation. Experiments report gains from SFT and further aesthetic refinement from RL.

7 Conclusion

  • The framework combines valid, challenging data synthesis, enhanced SFT supervision, and GRPO-based verifiable aesthetic rewards to improve presentation generation.The reported experiments find improvements for dense and MoE models, with RL refining multiple aesthetic dimensions beyond text-only supervision.Source: §7 · p. 10

AI Use Statement

The authors disclose using generative AI for synthetic-data generation, dataset cleaning and reformatting, and translation, and state that they reviewed and verified the resulting work.

AI Use Statement

  • Generative AI was used for synthetic datasets, dataset cleaning and reformatting, and translation; the authors reviewed the assisted work and take responsibility for the final content.They report no use for other tasks requiring disclosure.Source: AI Use Statement · p. 10

A Details of Presentation Generation Method Selection

The appendix compares code-based and image-based presentation generation, then evaluates HTML and Python configurations within OpenClaw. Single-HTML performs best overall for both tested models, while HTML has visual-design advantages over Python.

A Details of Presentation Generation Method Selection

  • Presentation generation approaches include Python code, HTML rendering, and direct image generation, which differ in flexibility, controllability, and extensibility.The study focuses on code-centric generation because image-based outputs can have textual hallucinations, instability, limited controllability, and higher computational costs.Source: A Details of Presentation Generation… · p. 12
  • OpenClaw is selected for its closed-loop planning and iteration, broad toolset, and support for multiple generation paradigms without framework modifications.This enables consistent evaluation across the paradigms.Source: A Details of Presentation Generation… · p. 12
  • The evaluation compares four configurations: single-page and multi-page HTML, and single-page and multi-page Python generation.Experiments use two models and evaluate on PresentBench and the GDPVal presentation-generation subset.Source: A Details of Presentation Generation… · p. 13
  • Single-HTML attains the best overall performance for both Qwen-3.6-27B and Qwen-3.6-35B-A3B across the evaluated generation settings.HTML-based generation also outperforms Python-based generation on visual design, including layout, color, diversity, and mathematical rendering.Source: Table 4 · p. 13, Table 5 · p. 13, A Details of Presentation Generation… · p. 14
  • Single-HTML is favored over Multi-HTML for robustness and optimization efficiency, with slide-wise generation supporting fault isolation and incremental refinement.Multi-HTML more readily introduces blank pages and rendering artifacts; the authors adopt Single-HTML for subsequent training.Source: A Details of Presentation Generation… · p. 14

B.1 Data Synthesis Pipeline

The section describes a cross-validated pipeline for synthesizing presentation tasks and rubrics, defining a broad domain taxonomy, and guiding supervised fine-tuning prompts, with attention to source fidelity, evaluation validity, and visual verification. It also specifies slide-deck deliverables and the hardware, optimization, and parallelism configuration for SFT runs.

B.1 Data Synthesis Pipeline

  • Task and rubric construction separates task scope from withheld answers, cross-checks completeness and correctness against source evidence, audits the pair, and tightens criteria without changing rubric structure.Rubric items cover source quantities and factual traps; tightening preserves dimensions, weights, and item counts.Source: §B.1 · pp. 14–15
  • The five-stage pipeline makes answer non-disclosure, source-grounded coverage, and evaluation difficulty explicit objectives for scalable task and rubric synthesis without human-in-the-loop authoring.Its stages are source profiling, task construction, rubric generation, principle audit, and rubric tightening.Source: §B.1 · pp. 14–15

B.2 Domain Taxonomy

  • The unified taxonomy used for data synthesis covers 25 top-level domains and 162 subdomains.Source: §B.2 · p. 15

C Details of Guided Supervised Fine-Tuning

  • Three prompt templates provide increasing guidance: base adds only task and output format, pro adds fidelity and layout checklists, and max prescribes a methodology with a render–inspect–fix loop.Each template is prepended to task instructions and followed by a fixed Expected Output block.Source: C Details of Guided Supervised… · p. 17
  • The prompt requirements prioritize source fidelity through traceable claims and exact numbers, while requiring slides to be rendered, inspected, and fixed until visual issues are resolved.The specified checks include overlap, overflow, small fonts, empty or unbalanced regions, and footer collisions.Source: C Details of Guided Supervised… · pp. 19–20
  • The expected output is a complete slide deck that satisfies the stated constraints, with one HTML file per slide in an html_slides folder.The examples given are slide_1.html, slide_2.html, and so on.Source: C Details of Guided Supervised… · p. 20

D.1 Training Details

  • SFT runs use full-parameter bf16 fine-tuning with thinking mode on a Megatron-based stack across 64 GPUs.Optimization uses Adam, a peak learning rate of 5e-6, cosine decay, weight decay 0.01, gradient clipping 1.0, and maximum context length 131,072 tokens.Source: §D.1 · p. 20
  • The Qwen-3.6-27B uses tensor-parallel = 2, context-parallel = 8, and pipeline-parallel = 4; Qwen-3.6-35B-A3B uses tensor-parallel = 2, context-parallel = 4, pipeline-parallel = 4, and expert-parallel = 2.Both layouts yield data-parallel = 1 across the 64 GPUs.Source: §D.1 · p. 20

D.2 Supplementary Results for SFT

This section identifies the scope of supplementary SFT results and introduces verifiable visual rewards, computed from fully rendered slides, to address visual flaws that imitation-based supervision does not explicitly optimize.

D.2 Supplementary Results for SFT

  • Table 7 reports per-module PresentBench scores for two target models under different teacher references and supervised fine-tuning configurations.The module abbreviations cover Academia, Advertising, Economics, Education, and Talk.Source: §D.2 · p. 20, Table 7 · p. 21

E Reward Design Details

  • SFT aligns presentations with teacher trajectories but does not explicitly optimize visual criteria tied to users’ first impressions.Post-SFT examples still show overflow, excessive whitespace, visual imbalance, and element overlap, which reduce readability and perceived quality.Source: E Reward Design Details · p. 21
  • Aesthetic quality is decomposed into five deterministic, interpretable rule-based functions over rendered slides rather than a learned critic or VLM judge.The final slide-level reward is an unweighted mean, and the deck-level reward averages across slides.Source: E Reward Design Details · p. 21

E.1 Render Infrastructure

  • The rendering service provides adaptive-height renders, Playwright screenshots, or semantic bounding-box trees, depending on the metric.The trees include element boxes, text-line rectangles, and resolved style attributes after layout and font resolution.Source: §E.1 · p. 22

E.2 Verifiable Aesthetic Metrics

The section defines five normalized, rule-based aesthetic rewards for recurring slide failures, then describes the rendering and optimization setup used to compute and train against them.

E.2 Verifiable Aesthetic Metrics

  • Five normalized metrics target aspect-ratio distortion, excessive whitespace, element collision, visual imbalance, and poor font readability; larger scores are better.The metrics are averaged without manual weighting to form slide-level and deck-level rewards.Source: §E.2 · pp. 22, 25
  • Aspect-ratio compliance compares responsive rendering against the 1280/720 target, tolerating mild deviations but penalizing vertically elongated layouts more heavily because they usually indicate overflow.The implementation uses a small margin and a sharper penalty for slides that become too tall.Source: §E.2 · pp. 22–23
  • Whitespace adequacy estimates local image variance from a grayscale screenshot, thresholds the result into a foreground map, and computes content ratio within a central region.The reward is mapped smoothly; nearly blank slides score low, while content-rich slides score high.Source: §E.2 · pp. 23–24
  • Collision avoidance counts qualifying overlaps between unrelated semantic elements and boundary violations, then decreases reward linearly with the total errors.A rectangle near the viewport edge is a boundary violation when its distance is below 8 px.Source: §E.2 · p. 24
  • Visual balance measures weighted content distribution across slide halves, excluding title- and footer-like blocks and tolerating common asymmetric layouts.Center-alignment factors and a mild power penalty make the metric tolerant rather than overly strict.Source: §E.2 · p. 24
  • Font readability scores text by size, exempts very small footnote-like regions, disables large-font penalties on title slides, and penalizes tiny or dense small-font text.Slides without text receive a neutral score of 0.5.Source: §E.2 · p. 25

F.1 Experimental Setup

  • The two backbones share the reward formulation, optimization objective, and stabilization strategy, while differing primarily in parallelization layout and scale.GRPO training uses a group-mean baseline, batch-level standard-deviation normalization, token-mean policy loss, asymmetric clipping, and a controlled KL-anchor comparison.Source: §F.1 · p. 25, Table 10 · p. 26, Table 11 · p. 26

F.2 Training Strategy Design

Human evaluation reports post-RL preference wins for both backbones, alongside supplementary PresentBench results across five subsets and five dimensions. Future work identifies gaps in content rewards, long-horizon credit assignment, and personalization.

F.2 Training Strategy Design

  • Each optimization step uses 64 rollouts from eight prompts, and matched experiments test whether an SFT-anchored KL term affects training stability across two backbones.Rollouts share prompt and seed within groups; the backbones are a 27B dense model and a 35B-A3B MoE model.Source: §F.2 · pp. 26–27
  • Table 12 lists the render infrastructure and verifiable aesthetic metrics used in reinforcement learning.Source: Table 12 · p. 27

F.3 Supplementary Results for RL

  • Post-RL results cover five PresentBench domain subsets and score rates across its five evaluation dimensions.Tables 7 and 8 report these supplementary results.Source: §F.3 · p. 27
  • 77.1% and 66.9% win rates for the dense backbone, and 82.4% and 71.8% for the MoE backbone, favor post-RL presentations in visual and overall quality, respectively.The blind comparison used three annotators and 150 cases per backbone.Source: §F.3 · p. 27

G Future Work

  • Content quality remains inherited largely from SFT rather than directly optimized during RL; proposed content rewards would assess faithfulness, completeness, and slide-level semantic alignment.The passage suggests manually scored cases as one route to training content reward models.Source: G Future Work · p. 27
  • Trajectory-level reward assignment remains coarse for long-horizon presentation generation, motivating step-level or hierarchical credit assignment and on-policy distillation.These alternatives may provide denser supervision and more context-aware attribution of visual outcomes to actions.Source: G Future Work · p. 28
  • The framework optimizes broadly preferred quality despite variation across users, domains, and scenarios; personalization would condition supervision and optimization on specific preferences.Proposed mechanisms include preference-conditioned fine-tuning and reward-conditioned reinforcement learning.Source: G Future Work · p. 28

Related papers

Loading 2610.30888v1…