Pith. sign in

REVIEW 3 major objections 9 references

Designing each slide in a high-level visual plan before writing code yields more coherent, preferred presentations than templates or direct code generation.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

T0 review

2026-07-13 08:51 UTC pith:RNXFQJKT

load-bearing objection Solid systems package for template-free slides with real human preference wins; the “design first” causal story is not cleanly isolated from pipeline depth. the 3 major comments →

arxiv 2605.26451 v1 pith:RNXFQJKT submitted 2026-04-08 cs.HC cs.CV

Design First, Code Later: Aesthetically Pleasing Template-Free Slides Generation

Zhiyao Cui , Chenxu Wang , Shuyue Hu , Yiqun Zhang , Wenqi Shao , Qiaosheng Zhang , Zhen Wang This is my paper
classification cs.HC cs.CV
keywords slide generationtemplate-free designhierarchical workflowmulti-agent reinforcement learningpresentation designmultimodal agentsPPTX generationdesign-implementation decoupling
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

Automatic slide tools usually either lock layouts into fixed templates or jump straight to executable code, so language models never get a real page-design stage. This paper claims that a hierarchical, template-free workflow—global style first, then per-page background, layout, and content design, then code—lets models compose freely while still producing stable, editable PPTX. A paired design–code dataset and multi-agent training support separate designer and coder models. The practical stake is audience attention: decks that win on layout, hierarchy, and color are the ones people actually prefer, and the reported human rankings put this approach ahead of both open pipelines and commercial systems in the study.

Core claim

DeepSlides establishes that decoupling slide-page design from implementation inside a hierarchical, template-free pipeline improves visual layout, hierarchy, and color and raises human preference relative to template-driven and direct code/HTML baselines. Design lives in a semantic specification; a coder module only then compiles that plan into runnable Python/PPTX, with iterative checks on completeness, compliance, and aesthetics. Supporting that claim, the authors release SlideDesign and train SlideQwens designer–coder pairs with multi-agent reinforcement learning.

What carries the argument

DeepSlides: a hierarchical design-first workflow that separates slides-level style and page-level design (background, layout, content) from code realization, optimized by multi-agent RL between a designer and a coder.

Load-bearing premise

The quality gains are mainly caused by decoupling design from code, not by extra retrieval, multi-step prompting, more refinement stages, or judges that reward the same aesthetics the system was trained on.

What would settle it

Match total compute, retrieval, and refinement steps between a design-first pipeline and a strong direct-code baseline, then re-run the same blind human ranking and layout/hierarchy scores; if the preference gap disappears, the decoupling claim fails.

Watch this falsifier. Get emailed when new claim-graph text bears on it.

Share X Bluesky LinkedIn Reddit HN

If this is right

  • Template-free hierarchical design can outrank both open academic pipelines and commercial tools on human Top-1 preference in comparable head-to-head tests.
  • Separating designer and coder roles reduces error propagation and lets aesthetics and executability be optimized on different rewards.
  • Automatically generated design–code pairs are enough to train small specialized models that emit runnable PPTX from page content and style constraints.
  • Completeness, compliance, and aesthetics feedback can drive both iterative page revision and multi-agent RL without fixed layout templates.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The same design-spec-then-render split likely helps other structured multimodal authoring tasks (posters, reports, UI mockups) where layout intent is currently fused with rendering code.
  • If automatic judges and training rewards share the same aesthetic rubric, part of the reported lift may be judge-alignment; hold-out human-only aesthetic tests would isolate that.
  • A portable design specification could become an interchange layer so one visual plan drives PPTX, HTML, or native app renderers.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Referee Report

3 major / 0 minor

Summary. The paper proposes DeepSlides, a hierarchical, template-free slide-generation workflow that first produces slides-level style and page-level design specifications (background, layout, content layers) and only then compiles them into editable PPTX via Python, with iterative completeness/compliance/aesthetics feedback. It also releases SlideDesign (topics under a Field-of-Science taxonomy with paired design/code resources) and trains SlideQwens designer/coder models via SFT plus multi-agent RL (rewards in Eqs. 1–3). Empirically, with Claude Haiku as the main backbone, DeepSlides scores highest on average VLM visual metrics in Table 2 and wins human Top-1 preference against open-source and commercial baselines (Figs. 8–9), with additional style ratings (Fig. 10) and small-model training gains (Table 3).

Significance. Automated slide generation is a practically important structured multimodal problem where current systems are either template-bound or code-first and often aesthetically weak. A design-first, template-free pipeline that still emits editable PPTX is a clear systems contribution; releasing code and a multi-disciplinary dataset, plus showing that a 0.6B backbone can be specialized for design and coding, are concrete assets for the community. Human preference results (anonymized order, open-source and commercial comparators) are stronger evidence of perceived quality than VLM scores alone. If the claimed benefit of decoupling design from implementation can be cleanly isolated, the paper would offer a reusable architectural pattern for other layout-constrained generation tasks, not only a stronger end-to-end demo.

major comments (3)
  1. The paper’s central framing (§1, §3, research question (2) in §5) attributes gains to decoupling design from implementation, but the experiments evaluate the full DeepSlides stack (DeepResearch/outline, web content expansion, slides-level style, three-layer design, iterative checker loop, then coding) against Direct, Auto-Slides, and EvoPresent (Table 2; Figs. 8–10). Shared research reports partially align inputs (§5.1), yet page-level expansion, multi-layer design, and design–eval iteration remain DeepSlides-specific. There is no controlled ablation that holds content, retrieval, and iteration budget fixed while only toggling an intermediate design specification versus joint design+code emission. Table 3 (SFT vs SFT+MARL) improves small-model executability/visual scores but does not test decoupling. Without that isolation, the causal claim for “design first, code later” is not establish
  2. Evaluator alignment weakens the independence of the aesthetic evidence. The MARL designer reward RD (§4.2, Eqs. 2–3) combines completeness, compliance, and aesthetics gated by implementation completeness from a VLM reward model; the main automatic evaluation (§5.2, Table 2) again uses a VLM judge on clarity/coherence/layout/hierarchy/color—closely related axes. Human preference (Figs. 8–10) is a better external signal, but the paper still leans on VLM scores to claim superiority on layout/hierarchy/color. Primary claims about visual quality should be anchored in human ratings (with uncertainty), with VLM metrics clearly secondary, or an independent judge/protocol should be shown not to share the training reward criteria.
  3. Human evaluation is directionally strong but under-specified for the breadth of the claim. Twenty volunteers rank/score 40 topics (20 per base model) with anonymized order (Figs. 8–10; Appendix D), yet the manuscript does not report inter-rater agreement, confidence intervals on win rates, or topic-level variance, nor how commercial systems (Kimi, Manus) were prompted/configured for fair topic-matched comparison. Given that Top-1 share and pairwise wins are the main support for real-world superiority, these statistics and protocol details are load-bearing for the preference claim.

Circularity Check

0 steps flagged

No circular derivation: DeepSlides is an empirical systems claim evaluated against external baselines and human raters, not a result forced by its own definitions or fits.

full rationale

This paper does not present a mathematical derivation chain in which a claimed prediction reduces to its inputs by construction. The load-bearing claims are empirical: a hierarchical, template-free design-then-code workflow (DeepSlides), a constructed dataset (SlideDesign), and MARL-trained SlideQwens models outperform baselines on objective render metrics, VLM-as-judge scores, and human preference. Design rewards (Eqs. 1–3) and VLM judges share aesthetic language, but the main Table 2 / human studies run Claude-backed DeepSlides against external systems without using the trained reward as the reported outcome, so automatic wins are not forced by the training objective. Human Top-1 preference (20 volunteers, anonymized ranking) is an independent external signal. Self-citation of ReMA is methodological inspiration only, not a uniqueness theorem that forbids alternatives. Training data filtered for executable code does not make baseline comparisons tautological. Causal confounding (pipeline depth vs. decoupling) is a validity concern, not circularity under the stated patterns. Score 0.

Axiom & Free-Parameter Ledger

3 free parameters · 4 axioms · 3 invented entities

This is an empirical systems paper. Load-bearing content is engineering choices and evaluation proxies, not physical postulates. Free parameters are training and reward weights; axioms are standard ML/HCI assumptions about what “good slides” mean and that VLM/human scores track quality; invented entities are named system modules and the dataset, which are constructed artifacts rather than unobserved natural kinds.

free parameters (3)
  • Designer reward weights α, β, γ
    RD = α r_cmp + β r_cpl + γ r_sty (Eq. 3); scalar weights chosen by authors and not derived from first principles; they shape what the designer optimizes.
  • SFT/MARL hyperparameters
    Learning rates (1e-4 SFT, 1e-6 GRPO), batch sizes, n=10 rollouts, 10 epochs, switch-agent every 10 steps, etc., are hand-set training knobs that affect reported SlideQwens success rates.
  • Balance metric mixture weight λ and Gaussian σ
    Objective Balance score mixes background-relative deviation and local contrast with λ and blur scale; definition is author-chosen and influences the Balance column in tables.
axioms (4)
  • ad hoc to paper Decoupling high-level design from executable PPTX code improves aesthetic quality without harming content quality relative to joint or template pipelines.
    Core methodological premise of DeepSlides (Abstract, §1, §3); supported empirically but not proven as a general law.
  • domain assumption Completeness, compliance, and aesthetics scores from a VLM reward/judge are valid enough proxies to train and rank slide quality.
    Used in page-level evaluation (§3.3) and MARL rewards (§4.2); human study partially validates but does not fully calibrate the proxy.
  • domain assumption Standard supervised fine-tuning plus multi-agent RL (ReMA-style alternating designer/coder) can specialize small LLMs for structured multimodal generation.
    Training paradigm assumption (§4); common in agent literature but outcome-dependent.
  • domain assumption Python-pptx editable PPTX is the appropriate target representation versus HTML/LaTeX for real-world slide workflows.
    Stated design choice in §3.3 Implementation and Table 1 comparison framing.
invented entities (3)
  • DeepSlides hierarchical workflow (slides-level style + three-layer page design + coder + checker loop) no independent evidence
    purpose: Organize template-free generation and separate aesthetic design from code realization.
    Primary system contribution; existence is by construction of the pipeline, not an external natural entity.
  • SlideDesign dataset independent evidence
    purpose: Provide paired content, design specs, and executable code for SFT/RL on slide tasks across FOS topics.
    New resource claimed in abstract and §4.1; independent value depends on release quality and licensing of sources.
  • SlideQwens (Designer and Coder models) no independent evidence
    purpose: Lightweight specialized models for layout design and PPTX code generation.
    Trained artifacts; performance claims rest on authors’ eval protocol.

reviewed 2026-07-13 · how reviews work

0 comments
Cite this review

Pith. "Pith review of Design First, Code Later: Aesthetically Pleasing Template-Free Slides Generation." pith.science (2026). https://pith.science/paper/RNXFQJKT

@misc{pith2026260526451,
  author       = {Pith},
  title        = {Pith review of: Design First, Code Later: Aesthetically Pleasing Template-Free Slides Generation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/RNXFQJKT}},
  note         = {Machine review of arXiv:2605.26451}
}
Share X Bluesky LinkedIn Reddit HN
read the original abstract

Producing presentation slides automatically entails coordinating narrative structure with page-level graphic design under strict spatial constraints. For such structured multimodal tasks, a well-organized design process is essential to ensure the final quality of slides. Existing approaches rely on fixed templates or directly emit executable code, thereby both limiting the creative layout-design capabilities of LLMs and bypassing the essential slide-page design step. To address these limitations, this paper (1) proposes a hierarchical slides generation workflow, DeepSlides, that systematically organizes slide design tasks without any predefined template or style, decoupling slide-page design from implementation; (2) introduces SlideDesign, a dataset tailored specifically for slides generation tasks; and (3) presents a multi-agent reinforcement learning training paradigm and trains a couple of models, SlideQwens, for slide design and implementation. Experimental results demonstrate that our proposed framework outperforms baseline methods on evaluated metrics and achieves superior performance in human preference evaluations. The dataset and code are available at https://github.com/sxswz213/DeepSlides.

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

9 extracted references · 2 linked inside Pith

  1. [1]

    Presenting a paper is an art: Self-improvement aesthetic agents for academic presentations.arXiv preprint arXiv:2510.05571. Manus. 2025. Manus. Accessed: 2025-12-02. Fanqing Meng, Lingxiao Du, Zongkai Liu, Zhixiang Zhou, Quanfeng Lu, Daocheng Fu, Tiancheng Han, Botian Shi, Wenhai Wang, Junjun He, Kaipeng Zhang, Ping Luo, Yu Qiao, Qiaosheng Zhang, and Wenq...

    arXiv 2025

  2. [2]

    Ziyu Wan, Yunxiang Li, Xiaoyu Wen, Yan Song, Hanjing Wang, Linyi Yang, Mark Schmidt, Jun Wang, Weinan Zhang, Shuyue Hu, and 1 others

    Slidecoder: Layout-aware rag-enhanced hi- erarchical slide generation from design.Preprint, arXiv:2506.07964. Ziyu Wan, Yunxiang Li, Xiaoyu Wen, Yan Song, Hanjing Wang, Linyi Yang, Mark Schmidt, Jun Wang, Weinan Zhang, Shuyue Hu, and 1 others

    Pith/arXiv arXiv

  3. [3]

    Xin Wang, Zhiyao Cui, Hao Li, Ya Zeng, Chenxu Wang, Ruiqi Song, Yihang Chen, Kun Shao, Qiaosheng Zhang, Jinzhuo Liu, Siyue Ren, Shuyue Hu, and Zhen Wang

    Rema: Learning to meta-think for llms with multi-agent reinforcement learning.arXiv preprint arXiv:2503.09501. Xin Wang, Zhiyao Cui, Hao Li, Ya Zeng, Chenxu Wang, Ruiqi Song, Yihang Chen, Kun Shao, Qiaosheng Zhang, Jinzhuo Liu, Siyue Ren, Shuyue Hu, and Zhen Wang. 2025. Perpilot: Personalizing vlm-based mobile agents via memory and exploration.Preprint, a...

    Pith/arXiv arXiv 2025

  4. [4]

    Import the necessary libraries

  5. [5]

    Create the slides and ensure the widescreen standard aspect ratio: 16:9 (13.33 inches × 7.5 inches)

  6. [6]

    According to the detailed description, add the title, bullet points, and images at specified positions; set fonts and styles; ex- plicitly set the size of each element to pre- vent overlap/occlusion; ensure text wraps automatically

  7. [7]

    Do not reserve any positions for any images that are not provided, and do not use text descriptions to fill the gaps

    Only the provided image URLs can be used. Do not reserve any positions for any images that are not provided, and do not use text descriptions to fill the gaps. Or you can also manually create some flowcharts using various graphics, but don’t just leave an empty space or just provide a textual description

  8. [8]

    All the text should be placed on the top layer

  9. [9]

    Note: output Python code only, do not out- put any other text

    Save the file as: <Save target> Previous code and errors (if any): <previous code> <error message> Please provide complete, executable Python code based on this information. Note: output Python code only, do not out- put any other text. Code will be save in utf-8 encoding. B Dataset We allocate 10 topics per secondary field (420 top- ics total). Our relea...

This paper was first reviewed by grok-4.5 on July 13, 2026.