EvoPresent Benchmark Overview

• EvoPresent Benchmark is a dual-purpose framework that evaluates both presentation generation quality and slide aesthetic awareness using annotated academic data.
• It employs objective metrics like perplexity, ROUGE-L, layout balance, and aesthetic scores to assess narrative coherence, content fidelity, and visual design quality.
• The benchmark supports iterative self-improvement in automated presentation systems by providing controlled aesthetic feedback and defect adjustment data.

Searching arXiv for the EvoPresent paper to ground the article in the source paper. EvoPresent Benchmark is a dual-purpose benchmark for academic presentation generation and slide aesthetic awareness, introduced together with the EvoPresent framework and the PresAesth multi-task reinforcement learning model in "Presenting a Paper is an Art: Self-Improvement Aesthetic Agents for Academic Presentations" (Liu et al., 7 Oct 2025). It is designed to support two tightly coupled objectives: evaluating how well automated systems transform research papers into coherent and visually effective presentation artifacts, and providing labeled aesthetic data for the scoring, defect adjustment, and comparison capabilities needed for iterative self-improvement. The benchmark comprises two components—Presentation Generation Quality and Aesthetic Awareness—and is positioned as the evaluation substrate that makes it possible to assess both presentation content construction and visual design quality within a single framework.

1. Scope, motivation, and conceptual position

EvoPresent Benchmark is motivated by the claim that existing automated presentation methods are difficult to evaluate rigorously because they often emphasize output appearance or text fidelity in isolation, while lacking reliable measures for narrative coherence, visual design quality, and especially aesthetic feedback and correction (Liu et al., 7 Oct 2025). In that framing, the benchmark is not only an external test suite; it is the infrastructure required to “evaluate it right” before self-improving agents can reliably refine their own outputs.

The benchmark has two major purposes. First, it evaluates academic paper presentation systems in a way that covers both the content of a presentation and its visual design. Second, it provides labeled aesthetic data for training and evaluating the model that powers self-correction in EvoPresent. This gives the benchmark a hybrid role as both an evaluation framework and a training substrate.

The benchmark is explicitly contrasted with prior systems such as PPTAgent, PresentAgent, Paper2Poster, and P2P. The stated distinction is breadth and realism: EvoPresent Benchmark spans multiple AI conference domains, includes multiple presentation formats, and introduces human-annotated aesthetic preference data with controlled design perturbations. A plausible implication is that the benchmark treats presentation generation as a multimodal, multi-criterion synthesis problem rather than as slide rendering alone.

2. Presentation Generation Quality

The Presentation Generation Quality component is built from 650 top-tier AI conference papers drawn from major venues such as ICLR and NeurIPS, spanning domains such as computer vision (CV) and NLP (Liu et al., 7 Oct 2025). Each paper is paired with multimodal presentation resources—slides, videos, and scripts—and these materials are annotated by 2–3 experts. The benchmark therefore evaluates systems against a richer target than a static deck: it asks whether a system can construct a coherent and realistic presentation package grounded in a scientific paper.

Evaluation is organized at two levels. The global evaluation layer uses objective metrics: Perplexity (PPL) for narrative coherence and fluency, ROUGE-L for content fidelity, Layout Balance for visual composition, and an Aesthetic score from PresAesth on a 1–10 scale. The fine-grained evaluation layer uses a VLM-as-judge on a 1–5 scale across eight dimensions. The content dimensions are Fidelity, Clarity, Narrative, and Engagement. The design dimensions are Elements, Layout, Hierarchy, and Color. This structure operationalizes the claim that strong presentation generation requires both semantic preservation and competent visual organization.

The source data also encode substantial stylistic variability. The benchmark reports that source papers average 9 pages, presentation videos average 9.5 minutes, slide counts range from 6 to 19, slide display times range from 10 to 100 seconds per slide, and script lengths range from 50 to 300 words. This suggests that the benchmark does not reduce academic presentation to a single canonical form, but instead reflects heterogeneous presentation practices across research domains.

3. Aesthetic Awareness

The Aesthetic Awareness component consists of 2,000 slide pairs constructed from real slide data via controlled perturbations (Liu et al., 7 Oct 2025). The construction process starts from original slides, applies perturbations such as style changes, creates three variants with different quality levels—poor, base, and good—and then forms slide pairs from those variants. Each slide is independently annotated by 2–3 annotators with quality ratings and defect labels.

The perturbation scheme is designed to create controlled differences in visual quality. In the appendix, the modifications are described as within-object alignment alterations, between-object layout alterations, and typography alterations. Examples given include changing text alignment or subcomponent positions, scaling and repositioning, spacing changes, and font size, weight, or spacing changes. The resulting score distribution is described as relatively balanced, with most samples concentrated around 5–7, and each slide contains at least 2–3 design errors across categories. This makes the task deliberately nontrivial.

This component supports three tasks. Scoring predicts an absolute aesthetic score for a single slide on a 1–10 scale. Defect adjustment identifies design deficiencies and describes improvements; the defect categories include Composition/Layout, Typography, Imagery/Visualizations, and No Deficiency. Comparison presents a baseline slide and two revisions, and requires choosing the better one. The benchmark is designed so that absolute judgments and comparative judgments reinforce one another: good scoring should assist defect diagnosis, and better diagnosis should support better comparison. The split is 1,600 training and 400 testing.

4. Evaluation protocol, metrics, and annotation

EvoPresent Benchmark is used in two ways: generation-quality evaluation over the 650-paper dataset and joint training–evaluation for aesthetic awareness over the labeled slide-pair dataset (Liu et al., 7 Oct 2025). For generation quality, the reported metrics are PPL with lower values preferred, ROUGE-L with higher values preferred, Layout Balance with higher values preferred, and the PresAesth aesthetic score with higher values preferred. Fine-grained VLM judging reports scores on the eight dimensions defined for content and design. For aesthetic awareness, the task metrics are MAE for scoring, F1-score for defect adjustment, and Accuracy for comparison.

The annotation pipeline is multi-layered. Generation-quality data are annotated by 2–3 experts. Aesthetic-awareness data use 2–3 annotators per slide. An overall human evaluation compares system outputs using five volunteers, and the appendix states that annotators were recruited from roughly 30 volunteers with design backgrounds for labeling. The annotation interface was implemented in Gradio.

The appendix also specifies a scoring protocol for slide aesthetics. Composition/Layout, Typography, and Imagery/Visualizations are each scored on a 1–10 scale. Annotators select specific deficiency types only when a dimension scores $\leq 4$. The rubric labels 9–10 as excellent, 7–8 as good, 5–6 as fair, 3–4 as poor, and 1–2 as very poor. This indicates that the benchmark encodes both continuous quality judgments and thresholded defect labeling.

The benchmark also anchors several computational definitions used in the paper. PPL is computed from a LLM conditioned on both the current and previous slide images; ROUGE-L is defined using LCS precision and recall; Layout Balance is defined via the distance from the visual center of mass to the slide center. For PresAesth, the reward is composed of a Format reward and an Accuracy reward, and the appendix gives a GRPO training objective in clipped-policy form with KL regularization, using $N = 8$ sampled responses per query, $\beta = 0.001$, $\alpha = 0.5$ for adjustment, and $\zeta = 0.25$ for scoring.

5. Role in EvoPresent and PresAesth

EvoPresent Benchmark is structurally tied to the EvoPresent system rather than merely accompanying it (Liu et al., 7 Oct 2025). EvoPresent is described as a four-agent draft–feedback–refinement pipeline consisting of a Storyline Agent, Scholar Agent, Design Agent, and Checker Agent. The Storyline Agent extracts structure and builds the script, the Scholar Agent enriches content, the Design Agent produces layout and visual rendering, and the Checker Agent evaluates output and gives feedback.

The Checker Agent depends on PresAesth, a multi-task RL aesthetic model that unifies scoring, defect adjustment, and comparison. The benchmark provides the training and evaluation data for these tasks, and PresAesth then supplies the stable, structured feedback used in EvoPresent’s iterative refinement loop. In this sense, the benchmark is the operational interface between aesthetic supervision and agentic self-improvement.

A common misconception would be to treat the benchmark as a presentation-generation leaderboard only. The paper instead presents it as the evaluation backbone of both the full-generation pipeline and the aesthetic feedback model. This suggests that the benchmark’s distinctive contribution lies in coupling end-task assessment with the internal judgment tasks required for self-correction.

6. Empirical findings, baselines, and limitations

The benchmark is used to compare multiple system families (Liu et al., 7 Oct 2025). For generation quality, baselines include Oracle; end-to-end models such as GPT-4o, GPT-5, Claude-4-Sonnet, DeepSeek-R1, and GPT-4o-Image; multi-agent systems such as PPTAgent-4o, PresentAgent-4o, and Paper2Poster-4o; and EvoPresent variants. For aesthetic awareness, PresAesth is compared with VLAA-Thinker-7B, Qwen2.5-VL-7B/32B, GLM-4.5V, Claude-4-Sonnet, GPT-4o, and GPT-5.

Several findings are emphasized. High-quality feedback is crucial: when PresAesth is used as the checker, EvoPresent’s aesthetic score rises from about 3.2 to over 8.0 within three iterations, whereas other checkers need more than five iterations and still perform worse. Initial capability does not equal correction ability: models with strong initial performance, including GPT-4o, DeepSeek-R1, Gemini-2.5-pro, and Claude-4-Sonnet, do not necessarily self-correct well in iterative feedback loops. There is a trade-off between content and design: reasoning-heavy models such as GPT-5 or DeepSeek-R1 can improve visual quality but sometimes introduce redundancy, increasing perplexity. Multi-task RL generalizes better: multi-task GRPO outperforms single-task RL and multi-task SFT across the aesthetic-awareness tasks.

The paper also acknowledges several limitations. Fine-grained evaluation still relies on VLM-as-judge. Aesthetic judgments are subjective, so labels use tolerance thresholds and human preference aggregation. The benchmark is centered on academic presentation formats, and generalization outside this domain is not guaranteed. Although multiple modalities are included, the core annotation focus remains on slide/image quality and narration-script alignment. These caveats do not negate the benchmark’s role; rather, they delimit the regime in which its measurements are intended to be interpreted.