AeSlides: Incentivizing Aesthetic Layout in LLM-Based Slide Generation via Verifiable Rewards

Pan, Yiming; Hu, Chengwei; Huang, Xuancheng; Huang, Can; Zhao, Mingming; Bi, Yuean; Zhang, Xiaohan; Zeng, Aohan; Hu, Linmei

AeSlides: Incentivizing Aesthetic Layout in LLM-Based Slide Generation via Verifiable Rewards

Yiming Pan^1,* Chengwei Hu² Xuancheng Huang^2,† Can Huang² Mingming Zhao²

Yuean Bi² Xiaohan Zhang² Aohan Zeng² Linmei Hu^1,†

¹ Beijing Institute of Technology ² Zhipu AI (Z.ai)

^* Work done during internship at Zhipu AI. ^† Corresponding authors.

Paper arXiv Code Model Dataset

AeSlides workflow overview and aesthetic deficiency categories — AeSlides targets aspect ratio, whitespace, collision, and visual imbalance for LLM-based slide generation through verifiable aesthetic rewards.

Abstract

Large language models (LLMs) have demonstrated strong potential in agentic tasks, particularly in slide generation. However, slide generation poses a fundamental challenge: the generation process is text-centric, whereas its quality is governed by visual aesthetics. This modality gap leads current models to frequently produce slides with aesthetically suboptimal layouts. Existing solutions typically rely either on heavy visual reflection, which incurs high inference cost yet yields limited gains; or on fine-tuning with large-scale datasets, which still provides weak and indirect aesthetic supervision. In contrast, the explicit use of aesthetic principles as supervision remains unexplored. In this work, we present AeSlides, a reinforcement learning framework with verifiable rewards for aesthetic layout supervision in slide generation. We introduce a suite of meticulously designed verifiable metrics to quantify slide layout quality, capturing key layout issues in an accurate, efficient, and low-cost manner. Leveraging these verifiable metrics, we develop a GRPO-based reinforcement learning method that directly optimizes slide generation models for aesthetically coherent layouts. With only 5K training prompts on GLM-4.7-Flash, AeSlides improves aspect ratio compliance from 36% to 85%, while reducing whitespace by 44%, element collisions by 43%, and visual imbalance by 28%. Human evaluation further shows a substantial improvement in overall quality, increasing scores from 3.31 to 3.56 (+7.6%), outperforming both model-based reward optimization and reflection-based agentic approaches, and even edging out Claude-Sonnet-4.5. These results demonstrate that such a verifiable aesthetic paradigm provides an efficient and scalable approach to aligning slide generation with human aesthetic preferences.

36% to 85% Aspect ratio compliance

-44% Excessive whitespace

-43% Element collision

-28% Visual imbalance

3.31 to 3.56 Human evaluation

Methodology

Modality Gap

Slide generators emit textual markup, while users judge the rendered visual layout. AeSlides treats this mismatch as the core bottleneck.

Render Infrastructure

A rendering pipeline collects screenshots, DOM structure, bounding boxes, and page geometry before metric computation.

Verifiable Metrics

Four layout issues are measured directly: distorted aspect ratio, excessive whitespace, element collision, and visual imbalance.

RL Optimization

GRPO training uses shaped multi-objective rewards, reward-decoupled normalization, and KL regularization for stable aesthetic alignment.

Main Experimental Results

AeSlides improves all four verifiable layout metrics and obtains the strongest human evaluation score among reported variants.

Proprietary models Base model Agentic reflection RL with model-based reward RL with verifiable rewards (Ours)

Group	Model	Render Error low	A.R. 1% high	A.R. 5% high	E.W. low	E.C. low	V.I. low	GPT-5-mini high	GPT-5.2 high	Human high

A.R.: aspect ratio compliance; E.W.: excessive whitespace; E.C.: element collision; V.I.: visual imbalance.

Meta Evaluation

Our verifiable metrics are more accurate, cheaper, and faster than VLM-based issue detection on AeSlides-Reward-Bench.

Metric Type	Distorted Aspect Ratio			Excessive Whitespace			Element Collision			Visual Imbalance			Latency low	Cost low
Metric Type	F1	F2	ROC-AUC	F1	F2	ROC-AUC	F1	F2	ROC-AUC	F1	F2	ROC-AUC	Latency low	Cost low
Coin-Flip Baseline	0.58	0.54	0.50	0.51	0.50	0.50	0.13	0.23	0.50	0.22	0.33	0.50	-	-
GPT-5.2	0.24	0.17	0.56	0.73	0.82	0.71	0.43	0.45	0.70	0.36	0.57	0.77	~8000ms	~$200
GPT-5-mini	0.54	0.43	0.68	0.41	0.31	0.61	0.19	0.13	0.55	0.47	0.56	0.75	~10000ms	~$30
GPT-5-nano	0.34	0.25	0.58	0.59	0.59	0.58	0.18	0.14	0.55	0.42	0.56	0.75	~12000ms	~$6
Verifiable Metrics (Ours)	0.83	0.91	0.77	0.80	0.88	0.89	0.47	0.51	0.79	0.61	0.77	0.83	~4000ms	~$0

Cost is estimated per 50K samples. Latency includes about 3000ms rendering time.

Case Gallery

Representative page-level and deck-level generations from the released AeSlides examples.

Page-Level Cases

Browse all page examples

Chinese agriculture risk management slide example

Deck-Level Cases

Browse all deck examples

Art Appreciation

Chinese cultural and visual-art presentation with consistent ink-painting style.

Platform Introduction

Product introduction deck with dashboards, diagrams, and system-overview slides.

API Documentation

Developer-facing documentation slides with structured concepts and endpoint-style explanations.

Sports Biography

Spanish-language narrative deck balancing imagery, timeline structure, and statistics.

Resources Click to expand

Model card

GLM-4.7-Flash-AeSlides

The released checkpoint fine-tunes GLM-4.7-Flash with verifiable aesthetic rewards for slide layout generation.

Base: GLM-4.7-Flash (30B-A3B)
Training: 5K prompts with GRPO-style RL
Result: 76% / 85% aspect ratio compliance; 3.561 human score

Open model card

Dataset card

AeSlides-Reward-Bench

Meta-evaluation data for slide aesthetics, annotated to stress-test verifiable metrics and VLM-based detectors.

Scope: 4 issue dimensions and 21 subcategories
Annotators: 6 slide-design annotators
Use: Metric reliability and reward-model comparison

Open dataset card

Appendix Insights and Diagnostics Click to expand

Theoretical Justification for GDPO

Standard GRPO normalizes the summed reward, which can let high-variance reward components dominate the advantage signal. AeSlides uses reward shaping and GDPO-style reward-decoupled normalization so aspect ratio, whitespace, collision, and visual imbalance each retain useful optimization signal.

Why KL Regularization Matters

Removing KL divergence can inflate verifiable rewards while collapsing policy entropy. The resulting policy tends toward overly conservative design patterns, which looks good under narrow metrics but weakens slide diversity and human preference.

Training Dynamics

Entropy training dynamics — Entropy analysis reveals collapse risk when KL is removed.

Rollout time dynamics — Verifiable rewards reduce reward-computation overhead compared with VLM rewards.

Human Evaluation and VLM Reliability

Bootstrap confidence intervals for human evaluation differences — AeSlides shows consistent human-score improvements over most baselines.

Pairwise win-rate heatmaps — Pairwise preferences support the aggregate human evaluation ranking.

Bradley-Terry model scores — Bradley-Terry scores provide another view of relative model quality.

Bland-Altman analysis of automatic versus human scores — Automatic VLM scores show limited agreement with human judgments.

Dataset Diagnostics

Filtered special page distribution — Structurally simple pages are filtered before RL training.

Language distribution in AeSlides-7k — AeSlides-7k covers multilingual slide generation settings.

Page index distribution in AeSlides-7k — Training samples cover varied positions within slide decks.

Prefix token count distribution — Long prefixes test context-conditioned slide generation.

Qualitative Appendix Cases

KL ablation qualitative case — KL ablation illustrates reward inflation with conservative layout collapse.

Failure cases of AeSlides — Failures observed suggest some aesthetic factors are not yet directly supervised.

End-to-end generation case one — End-to-end deck generation case study.

End-to-end generation case two — End-to-end multilingual and culture-specific generation case study.

BibTeX

@article{pan2026aeslides,
  title={AeSlides: Incentivizing Aesthetic Layout in LLM-Based Slide Generation via Verifiable Rewards},
  author={Pan, Yiming and Hu, Chengwei and Huang, Xuancheng and Huang, Can and Zhao, Mingming and Bi, Yuean and Zhang, Xiaohan and Zeng, Aohan and Hu, Linmei},
  journal={arXiv preprint arXiv:2604.22840},
  year={2026},
  doi={10.48550/arXiv.2604.22840},
  url={https://arxiv.org/abs/2604.22840}
}