---
title: 'EvoPresent: Automated Presentation Generator'
url: https://www.emergentmind.com/topics/evopresent
type: topic
---

# EvoPresent: Automated Presentation Generator

EvoPresent is a self-improvement agent framework for automated academic presentation generation that unifies coherent narratives, aesthetic-aware designs, and realistic presentation delivery via virtual characters. It was introduced to address three difficulties identified in existing automated methods: limited storytelling, insufficient aesthetic quality, and constrained self-adjustment. At the center of the framework is PresAesth, a multi-task reinforcement learning aesthetic model that provides aesthetic scoring, defect adjustment, and comparative feedback, enabling iterative refinement even under limited aesthetic training data. The accompanying EvoPresent Benchmark combines a presentation-generation track built on 650 top-tier AI conference papers with multimodal resources and an aesthetic-awareness track based on 2,000 slide pairs with varying aesthetic levels [2510.05571].

## 1. Motivation and problem formulation

EvoPresent is situated in the context of academic presentation automation for research dissemination. The framework is motivated by the observation that existing automated pipelines often produce presentations with weak narrative coherence, suboptimal visual design, and limited capacity for iterative correction. The paper summarizes this limitation through a simple principle: there is no way to improve it when evaluation is unreliable [2510.05571].

The technical problem is therefore not only generation but also evaluation-guided self-correction. In the formulation underlying EvoPresent, presentation quality spans both content and design. Content quality concerns storyline construction, script organization, and presentation flow. Design quality concerns layout, hierarchy, typography, imagery, and color. Self-improvement requires a checker that can score outputs, identify deficiencies, and compare alternatives in a way that is actionable for subsequent refinement.

This framing places EvoPresent in a broader class of agentic systems, but with a specific emphasis on academic papers as source material and on presentations as multimodal outputs. A plausible implication is that the framework treats presentation authoring as a closed-loop optimization problem rather than as a one-shot conversion from paper to slides.

## 2. Agentic architecture

EvoPresent adopts a draft–feedback–refinement iterative agentic loop in which four specialized agents interact at each round [2510.05571]. The architecture is organized around division of labor rather than a monolithic generator.

The **Storyline Agent** extracts and organizes core text and visuals from the research paper, constructs an initial storyline, and produces a complete script. The description specifies that it employs tools such as Marker for PDF extraction and segments content by themes, with the goals of ensuring coverage and minimizing redundancy.

The **Scholar Agent** enriches the storyline through knowledge enrichment and visual enhancement. The stated examples include searching citations, retrieving related work, and generating relevant figures with GPT-4o or Qwen-Image. In the architecture, this agent extends the initial narrative with contextual and visual material rather than redesigning the storyline from scratch.

The **Design Agent** converts the enhanced script into visual slides or video frames. It contains a **Layout Planner**, which calculates stable, balanced layouts for text and visuals, and a **Style Render** component, which applies a consistent style via CSS-like themes. EvoPresent uses HTML as the slide format, and the paper states that this choice offers fine-grained control over appearance and flexibility compared to PPTX.

The **Checker Agent** evaluates both design and content, provides improvement feedback, triggers further refinement cycles, and chooses the best candidate using PresAesth’s scores. The architecture therefore combines global assessment and local revision guidance within the same feedback path.

This organization reflects a recurring design principle in the framework: content construction, external knowledge enrichment, visual design, and evaluation are separated into explicit components. This suggests that EvoPresent is intended to reduce interference between narrative planning and aesthetic optimization, while still allowing the components to interact through iterative feedback.

## 3. PresAesth and aesthetic awareness

PresAesth is the core evaluation model in EvoPresent. It is described as a multi-task RL aesthetic model trained to perform three tasks: scoring, defect adjustment, and comparison [2510.05571].

In the **scoring** task, the model receives a slide image and assigns an overall aesthetic quality score as a floating-point value. In the **defect adjustment** task, it diagnoses concrete design deficiencies among composition/layout, typography, and imagery/visualizations, then provides targeted improvement advice. In the **comparison** task, it receives a baseline slide and two altered slides and determines which revision is superior in aesthetic quality.

PresAesth uses **Group Relative Policy Optimization (GRPO)** and is based on Qwen-2.5-VL-7B. The training signal combines a format reward and an accuracy reward. The accuracy reward is task-specific:

$$
r_\mathrm{acc} =
\begin{cases}
\mathbb{I}(o_\mathrm{comp} = y_\mathrm{comp}) & \text{(Comparison Task)} \\
\mathbb{I}(\mathrm{F1}(f(o_\mathrm{def}), y_\mathrm{def}) > \alpha) & \text{(Adjustment Task)} \\
\mathbb{I}(|o_\mathrm{score} - y_\mathrm{score}| < \zeta) & \text{(Scoring Task)}
\end{cases}
$$

The total reward is defined as:

$$
r^{(i)} = r_\mathrm{fmt}^{(i)} + r_\mathrm{acc}^{(i)}
$$

For a query \(q\), the normalized advantage is computed as:

$$
\hat{A}^{(i)} = \frac{r^{(i)} - \operatorname{mean}(r^{(1)}, \dots, r^{(N)})}{\operatorname{std}(r^{(1)}, \dots, r^{(N)})}
$$

and the GRPO objective is:

$$
\mathcal{J}(\theta) = \mathbb{E}_{q, o^{(i)} \sim \pi_{\theta_\textrm{old}}}
\left\{
\min\left[\rho^{(i)} \hat{A}^{(i)}, \operatorname{clip}(\rho^{(i)}, 1-\delta, 1+\delta)\hat{A}^{(i)}\right]
- \beta \cdot \mathbb{D}_{\textrm{KL}}[\pi_{\theta_\textrm{new}} || \pi_\textrm{ref}]
\right\}
$$

where

$$
\rho^{(i)} = \frac{\pi_{\theta_\textrm{new}}(o^{(i)}|q)}{\pi_{\theta_\textrm{old}}(o^{(i)}|q)}
$$

The paper also states that the dataset for PresAesth contains approximately 3,400 human-annotated slides, spanning academic and general material, with splits across comparison, defect, and scoring labels. Training details reported in the summary include initialization from Qwen-VL-7B, about 2 epochs on 8 × H100, single-image resizing to longest edge 960px, three-image comparison resizing to 720px, defect-task threshold \(\alpha=0.5\), and scoring tolerance \(\zeta=0.25\).

A common misconception in automated presentation generation is that aesthetic evaluation can be treated as a simple scalar preference model. PresAesth’s design instead assumes that useful evaluation must be multi-task: it must assign scores, localize defects, and choose between alternatives.

## 4. Iterative self-improvement mechanism

EvoPresent’s self-improvement mechanism is implemented through the Checker Agent, which uses PresAesth inside an iterative optimization loop [2510.05571]. The reported procedure is explicit: for each presentation draft, PresAesth scores current slides; if quality exceeds a threshold, the system outputs the result as final; otherwise it issues corrective feedback to the Layout Planner or Design Agent; if quality drops after revision, the system reverts to the previous best; and the process iterates up to \(T\) rounds or until satisfactory quality is reached.

This loop gives EvoPresent a revision policy rather than a mere reranking function. The paper emphasizes that the checker provides both global coherence judgments and local, actionable feedback. That distinction matters because aesthetic failure modes are not limited to single-slide appearance; they also arise from interactions between storyline structure and page-level design.

The reported behavior of the loop is that it often converges within three iterations when high-quality feedback from PresAesth is available. The framework is therefore presented as depending critically on the quality of its evaluator. The paper’s findings state that high-quality feedback is essential for agent self-improvement, while initial capability alone does not guarantee effective self-correction [2510.05571].

This finding directly counters the view that stronger base generators automatically yield better iterative refinement. EvoPresent’s experiments indicate that an initially strong model without targeted feedback can remain weak at correction, whereas a system with a better checker can improve substantially over several rounds.

## 5. EvoPresent Benchmark

The EvoPresent Benchmark is divided into two tracks: **Presentation Generation Quality** and **Aesthetic Awareness** [2510.05571]. The benchmark is designed to evaluate both content construction and design quality, as well as the evaluative abilities needed for self-improvement.

The **Presentation Generation Quality** track is built on 650 papers from top AI conferences, including resources such as slides, presentation scripts, and presentation videos. Annotations are provided by 2–3 experts per sample. The data statistics reported in the summary are an average of 9 slides and 9.5 minutes per presentation, with slide decks ranging from 6 to 19 pages.

The **Aesthetic Awareness** track consists of 2,000 slide pairs, partitioned into 1,600 training and 400 testing examples, constructed by controlled visual perturbations such as alignment, spacing, and typography. It supports three evaluation settings: scoring with MAE against human judgments, defect adjustment with F1 over the categories No Deficiency, Layout, Typography, and Imagery/Visualizations, and comparison with classification accuracy.

The benchmark’s reported global metrics for presentation generation include perplexity, ROUGE-L, layout balance, and aesthetic score. The formulas given in the summary are:

$$
\mathrm{PPL} = \exp\left( -\frac{1}{N} \sum_{i=1}^N \log P(w_i \mid I_\mathrm{curr}, I_\mathrm{prev}, w_1,\ldots,w_{i-1}) \right)
$$

$$
\mathrm{ROUGE\mbox{-}L} = \frac{(1+\beta^2)R_\mathrm{lcs} P_\mathrm{lcs}}{\beta^2 R_\mathrm{lcs} + P_\mathrm{lcs}}
$$

$$
\mathrm{Balance} = \max\left(0, 1 - \frac{d}{d_{\max}}\right)
$$

where \(d\) is the distance from the weighted center of mass to the slide center and \(d_{\max} = \sqrt{2}/2\). In addition, the benchmark includes fine-grained VLM-as-judge evaluation on eight axes: Content—Fidelity, Clarity, Narrative, Engagement; and Design—Elements, Layout, Hierarchy, Color.

## 6. Empirical findings, trade-offs, and significance

The paper reports that EvoPresent outperforms previous multi-agent and end-to-end generation systems, including PPTAgent, PresentAgent, and GPT-4o-based approaches, on numerous axes, with lower perplexity, higher ROUGE-L and balance, and stronger fine-grained VLM-as-judge scores, especially for design [2510.05571]. It also reports that PresAesth matches human preferences better than closed-source models such as GPT-4o and Claude-4-Sonnet in scoring, adjustment, and comparison tasks.

The experiments emphasize the role of feedback quality. With PresAesth as checker, the paper states that agents improve from mediocre aesthetics, reported as score 3.2, to excellent aesthetics, reported as above 8.0, within three rounds. Competing checkers produce slower and less pronounced improvement. Even models described as having high initial capability, such as Gemini-2.5-pro, are reported to show weak correction when they lack targeted, model-generated improvement signals.

A second major result is that **multi-task GRPO RL shows stronger generalization in aesthetic awareness tasks**. The ablation summary states that removing joint training or replacing RL with supervised fine-tuning impairs generalizability and actionable feedback. The paper characterizes this as evidence that surface-level label fitting is not sufficient for generating useful corrective guidance.

A third result is the existence of a trade-off between visual design and content construction. The paper states that automated generation pipelines exhibit such a trade-off; strong reasoning models tend to favor content but may clutter slides, while visual-leaning agents may improve aesthetics at the expense of narrative flow. EvoPresent is presented as improving this balance through its iterative, agentic, and aesthetic-aware design.

Human and simulated evaluations further indicate that EvoPresent videos rank highest on simulated human metrics such as fluency and comprehension, and that in side-by-side tests EvoPresent is favored over other automated methods and is competitive with human references. The ablations also report that removing the Scholar, Design, or Checker agent degrades content, design, or aesthetic scores substantially.

Taken together, these results position EvoPresent as a framework in which evaluation is not ancillary but constitutive. Its significance lies less in any single generator component than in the integration of narrative planning, design synthesis, and multi-task aesthetic feedback into a single iterative system for academic presentation generation [2510.05571].

Source: https://www.emergentmind.com/topics/evopresent