AeSlides: RL for Aesthetic Slide Layout

• AeSlides is a reinforcement learning framework that targets enhanced slide aesthetics by optimizing layout quality through verifiable reward signals.
• It explicitly quantifies key layout failures—distorted aspect ratio, excessive whitespace, element collision, and visual imbalance—to guide slide generation.
• State-of-the-art RL techniques in AeSlides yield significant improvements in slide quality, increasing aspect ratio compliance and reducing visual defects.

Searching arXiv for the AeSlides paper and closely related follow-up usage. arxiv_search(query="(Pan et al., 21 Apr 2026) OR AeSlides Incentivizing Aesthetic Layout in LLM-Based Slide Generation via Verifiable Rewards", max_results=5) arxiv_search(query="(Dong, 9 Jun 2026) OR Visual-SDPO AeSlides", max_results=5) AeSlides is a reinforcement learning framework with verifiable rewards for Aesthetic layout supervision in Slide generation. It addresses a core modality gap in LLM-based slide generation: the model generates structured text, specifically HTML, while slide quality is judged primarily in the visual domain. AeSlides therefore treats slide aesthetics as an optimization target rather than as an indirect byproduct of text supervision, and formalizes four recurrent layout failures—distorted aspect ratio, excessive whitespace, element collision, and visual imbalance—as explicit reward dimensions for post-training (Pan et al., 21 Apr 2026).

1. Task formulation and scope

AeSlides studies prefix-conditioned next-slide generation. At generation step $t$, given a user query $Q$, supporting context $C$, and preceding slides $P_{<t}$, the model generates the next slide $P_t$:

$P_t \sim G_\theta(Q, C, P_{<t}).$

The generated slide is represented as HTML, rendered through a browser-based stack built with Playwright, FastAPI, and Uvicorn. Reward computation is therefore based on the actual rendered result rather than on raw source text alone (Pan et al., 21 Apr 2026).

The framework does not attempt to solve the entire presentation-authoring problem. Its focus is narrower and more specific: improving the page-level visual layout quality of generated slides. The training setup excludes many structurally simple page types, including cover pages, tables of contents, dividers, and ending pages, because those follow different design principles and are already relatively easy. This makes AeSlides primarily a post-training method for dense content slides, where space allocation and compositional stability are most difficult (Pan et al., 21 Apr 2026).

A common misconception is to treat AeSlides as a general deck-planning or multimodal reflection system. Its contribution lies elsewhere. The framework does not center on iterative critique loops, nor on a new renderer, nor on a new content planner. Its central claim is that several aesthetically important slide-layout properties are structured enough to be programmatically verified, and that these verifiable signals can supervise RL more effectively than opaque learned reward models or heavy reflection pipelines (Pan et al., 21 Apr 2026).

2. Targeted layout failures and metric design

AeSlides concentrates on four high-impact layout issues. These are the dimensions the paper identifies as especially important in its benchmark analysis and reward design.

Issue	What it captures	Measurement basis
Distorted aspect ratio	Overly tall or otherwise non-standard slide shape	Dynamic rendered width/height
Excessive whitespace	Large underutilized empty regions	Local grayscale variance map
Element collision	Overlap, overflow, out-of-bound violations	DOM trees and bounding boxes
Visual imbalance	Uneven distribution of visual weight	Layout centroid distance

For distorted aspect ratio, AeSlides renders the page with an initial viewport of $1280 \times 10$, uses JavaScript to retrieve the maximum of scrollHeight and offsetHeight for both body and html, resets the viewport height accordingly, and then computes the final ratio from the rendered width and height. The target is the standard slide ratio $16:9$. The reward is shaped with an asymmetric quadratic form:

$$\mathcal{R}(x)=\exp\bigg( -\alpha\big( \log(\frac{x}{\text{target}}) \big)^2-\beta\cdot\max\big(-\log(\frac{x}{\text{target}})-m, 0 \big)^2 \bigg),$$

with $\alpha = 16.0$, $Q$0, and $Q$1. The asymmetry penalizes overly tall slides more heavily, because overlong pages are a frequent failure mode (Pan et al., 21 Apr 2026).

For excessive whitespace, the framework converts the rendered image to grayscale, computes a local variance map over a box-filter neighborhood of size $Q$2, $Q$3, clips the local standard deviation with $Q$4, applies threshold $Q$5, and removes boundary regions. The metric is intended to estimate visually unproductive blank space rather than semantic emptiness. The paper also documents reward-hacking behavior here: the model attempted to exploit the metric by inserting textured or image-based backgrounds, and the authors mitigated this by removing background images and smoothing with a Gaussian kernel of size $Q$6 before variance estimation. This suggests that one of the advantages of verifiable rewards is not perfection, but debuggability (Pan et al., 21 Apr 2026).

For element collision, AeSlides filters the DOM into perceptually meaningful visual units, excluding invisible nodes and overly fine-grained primitives such as SVG primitives, then detects three event types: overlap between unrelated elements, escape from parent containers, and overflow beyond slide boundaries. Certain benign cases, such as weak transparent backgrounds and corner badges, are excluded unless they occlude main content. The resulting collision score is monotonic, with higher values indicating worse layout (Pan et al., 21 Apr 2026).

For visual imbalance, AeSlides computes the visual centroid of the layout and measures its ellipse-normalized distance from the slide center:

$Q$7

with $Q$8 and $Q$9. Horizontal imbalance is treated as more perceptually salient than vertical imbalance. This metric is explicitly about compositional stability rather than overlap or density (Pan et al., 21 Apr 2026).

3. Reinforcement-learning formulation

AeSlides builds on Group Relative Policy Optimization (GRPO) and incorporates several later stabilizing refinements, including clip-higher and token-level loss from DAPO, sequence-level importance ratio from GSPO, rollout routing replay, and reward-decoupled normalization from GDPO. The final policy objective is:

$C$0

The rollout group size is $C$1, and reward is computed after rendering each sampled slide (Pan et al., 21 Apr 2026).

A key technical choice is reward-decoupled normalization. If $C$2 is reward component $C$3 for rollout $C$4 under prompt $C$5, each reward component is normalized independently across the rollout group:

$C$6

then summed:

$C$7

and batch-normalized to obtain $C$8. This prevents a high-variance reward from collapsing the contribution of lower-variance components. The paper identifies this as essential for multi-objective aesthetic optimization (Pan et al., 21 Apr 2026).

AeSlides also retains KL regularization, in contrast to some later RL practices that remove it. Here the reason is domain-specific. The base model is already competent at slide generation; the goal is to improve layout without collapsing semantic diversity. In the reported ablation, removing KL yields superficially excellent metric scores but drives the policy toward conservative template-like designs. This is an instance where raw reward maximization diverges from desirable generative behavior (Pan et al., 21 Apr 2026).

The backbone used is GLM-4.7-Flash (30B-A3B), with 29.94B total parameters and 3.90B activated parameters. Training uses 8×H100 GPUs, 1 epoch, learning rate $C$9, weight decay 0.1, group size 8, rollout max response length 8192, rollout temperature 1, KL coefficient 0.1, and clipping parameters $P_{<t}$0, $P_{<t}$1 (Pan et al., 21 Apr 2026).

4. Data resources and evaluation protocol

AeSlides introduces two data resources. AeSlides-Reward-Bench is used to meta-evaluate the reward metrics. It contains 2068 records, 1029 paired tasks, 10 singleton tasks, annotations from 6 trained annotators, and about 300 person-hours of annotation effort. It covers 4 issue categories and 21 issue subcategories (Pan et al., 21 Apr 2026).

The RL training and evaluation corpus is AeSlides-7k. The paper reports a deck-level split of 85:15. Appendix statistics give 8471 collected page-level samples, 1735 filtered special pages, and a remaining split of 5704 train and 1032 eval. The abstract simplifies this as training with only 5K training prompts. The dataset is multilingual, with more than five languages including English, Chinese, Arabic, Thai, and Cyrillic-script languages (Pan et al., 21 Apr 2026).

Metric validation compares AeSlides’ verifiable detectors to GPT-based issue detectors. For aspect ratio, the verifiable detector achieves F1 0.83, F2 0.91, ROC-AUC 0.77; for whitespace, F1 0.80, F2 0.88, ROC-AUC 0.89; for collision, F1 0.47, F2 0.51, ROC-AUC 0.79; and for visual imbalance, F1 0.61, F2 0.77, ROC-AUC 0.83. The same table reports latency of about 4000 ms end-to-end for verifiable metrics versus about 8000–12000 ms for GPT detectors, with near-zero cost for the former and nontrivial per-sample cost for the latter. This directly supports the paper’s argument that targeted verifiable signals can outperform model-based judges on the exact dimensions they supervise (Pan et al., 21 Apr 2026).

Automatic evaluation on AeSlides-7k-eval reports rendering error, aspect ratio compliance at 1% and 5% tolerance, whitespace, collision, visual imbalance, and VLM-based overall scores from GPT-5-mini and GPT-5.2. Human evaluation is also reported, with annotators experienced in slide design, anonymized model outputs, and approximately 70% of score differences within 1 point; the paper reports $P_{<t}$2 (Pan et al., 21 Apr 2026).

5. Empirical performance and comparative results

The central empirical result is that, 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 shows overall quality increasing from 3.31 to 3.56 (+7.6%) (Pan et al., 21 Apr 2026).

More specifically, the base GLM-4.7-Flash model reports 24% / 36% aspect ratio compliance at 1% / 5% tolerance, whitespace 0.059, collision 0.126, visual imbalance 1.545, and human score 3.314. AeSlides: Verifiable Rewards w/ GDPO reaches 76% / 85%, whitespace 0.033, collision 0.072, visual imbalance 1.111, and human score 3.561. The same table reports Claude-Sonnet-4.5 at human score 3.442 and GPT-5.2 at 3.076, with much weaker aspect-ratio compliance than AeSlides (Pan et al., 21 Apr 2026).

These results are notable because AeSlides is explicitly compared against three common alternatives. First, reflection-based refinement via DeepPresenter performs worse than the base model on several layout dimensions, with render error 1.74%, aspect ratio 27% / 41%, whitespace 0.070, collision 0.212, visual imbalance 2.170, and human score 3.107. The paper attributes this to multimodal reflection failing to detect subtle defects and to the high overhead of iterative rendering. Prompt tokens increase by 155% and completion tokens by 46% relative to single-pass generation (Pan et al., 21 Apr 2026).

Second, model-based reward optimization with GPT-5-nano or GPT-5-mini reward signals improves some metrics but remains weaker than verifiable rewards. The paper highlights a particularly important inconsistency: a visually rewarded GPT-5-mini variant receives a higher GPT-5-mini score than AeSlides, yet is worse on nearly all verifiable metrics and worse in human evaluation. AeSlides presents this as evidence of reward-model blind spots and self-reinforcing bias in learned judges (Pan et al., 21 Apr 2026).

Third, pure supervised fine-tuning is not enough. The base model already benefits from large-scale slide SFT, yet still shows severe layout defects. AeSlides therefore functions as a post-training stage that translates latent slide competence into explicit visual-layout optimization (Pan et al., 21 Apr 2026).

The ablation studies reinforce the causal role of the reward design. Removing the aspect-ratio reward drops aspect-ratio compliance from 85% to 39%; removing the whitespace reward worsens whitespace from 0.033 to 0.093; removing the collision reward worsens collision from 0.072 to 0.200; and removing the visual-imbalance reward worsens imbalance from 1.111 to 2.289. Removing KL produces near-maximal metric scores but low entropy (0.072 versus 0.099 in the full method) and conservative template-like behavior. This makes clear that AeSlides is not merely a collection of heuristics, but a carefully balanced multi-objective RL system (Pan et al., 21 Apr 2026).

6. Interpretation, limitations, and later use

AeSlides’ broader significance lies in reframing slide aesthetics as a verifiable RL problem. Its argument is not that hand-engineered metrics are universally superior to learned judges, but that in domains where quality depends on structured geometric and compositional properties, explicit verifiable supervision can be both more faithful and more efficient. This is especially persuasive here because the paper shows that GPT-based critics are weak precisely on the four issues the RL process needs to detect reliably (Pan et al., 21 Apr 2026).

The framework nonetheless has clear limitations. It supervises only four dimensions and leaves out many others, including insufficient content density, user-intent mismatch, suboptimal color schemes, unsuitable image selection, style consistency, and typography-related issues. The paper also notes that element collision and visual imbalance rely heavily on heuristics, which creates engineering overhead and possible brittleness. A further limitation is preference heterogeneity: the reward design assumes a relatively shared notion of aesthetics, whereas actual presentation preferences vary across users, domains, and cultures. The paper therefore suggests more personalized or adaptive aesthetic tuning as future work (Pan et al., 21 Apr 2026).

AeSlides has also become a benchmark substrate beyond its original RL formulation. In “Self-Distillation Policy Optimization via Visual Feedback: Bridging Code and Visual Artifacts”, the authors adopt the AeSlides-7k page-level prompts as the slide domain and use the AeSlides verifiable metrics—aspect ratio, whitespace, collision, imbalance—on the AeSlides-7k-eval split with decks held out. In that setting, the reported Avg score improves from 49.5 for zero-shot Qwen3-VL-8B-Instruct to 60.7 with full Visual-SDPO, indicating that AeSlides functions not only as a training framework but also as a reusable slide-quality benchmark for later visually grounded policy optimization work (Dong, 9 Jun 2026).

A plausible implication is that AeSlides marks a shift in slide-generation research from generic “better-looking outputs” toward auditable alignment targets. Rather than asking a model or evaluator to internalize all aesthetic judgment implicitly, it exposes a subset of slide-quality criteria as measurable, controllable objects. For slide generation, where small geometric failures can dominate user perception, that shift is methodologically consequential.