Skip to content

Presenting a Paper is an Art: Self-Improvement Aesthetic Agents for Academic Presentations

🕒 Published (v1): 2026-01-01 · Source: ICLR · Venue: ICLR 2026 · link

Ask a follow-up

Open an assistant pre-loaded with this paper's context.

💬 Ask ChatGPT✦ Ask Claude

TL;DR

EvoPresent is a self-improving multi-agent framework that converts academic papers into presentation videos (slides, scripts, video) via a draft–feedback–revision loop driven by a dedicated aesthetic reward model. The core enabler is PresAesth, a multi-task RL model trained with GRPO to score slides, detect defects, and compare revisions, replacing brittle VLM-as-judge evaluators. A new benchmark of 650 papers and 2,000 annotated slide pairs supports systematic evaluation.

Problem

Automated paper-to-presentation pipelines (PPTAgent, PresentAgent, Paper2Poster) suffer from three compounding failures: (1) they extract content directly without narrative restructuring, yielding low coherence; (2) they rely on fixed templates, limiting visual design flexibility; and (3) they lack reliable self-improvement loops—either using no checker or using unstable VLM-as-judge scoring—forcing heavy manual revision. Existing aesthetic evaluation benchmarks do not cover academic slide aesthetics.

Method

EvoPresent orchestrates four specialized agents in a sequential pipeline:

  1. Storyline Agent — uses Marker to extract text/figures, then constructs thematic chapters, associates visuals, and outputs a full presentation script.
  2. Scholar Agent — enriches content via ArXiv MCP tools and generates supplementary visuals with GPT-4o/Qwen-Image.
  3. Design Agent — a two-stage layout planner (initial geometry → CSS style render) that outputs HTML slides with dynamic effects.
  4. Checker Agent — runs Algorithm 1: iteratively calls PresAesth to score the current slide \(S^{(t)}\); if score \(\geq S_{th}=8.0\) it exits early; otherwise it generates targeted feedback, reverts to \(S^{(t-1)}\) if quality degraded, and feeds back to the Layout Planner.

PresAesth fine-tunes Qwen-2.5-VL-7B via GRPO across three joint tasks. The reward is \(r^{(i)} = r_{\text{fmt}}^{(i)} + r_{\text{acc}}^{(i)}\) where: $\(r_{\text{acc}} = \begin{cases} \mathbb{I}(o_{\text{comp}} = y_{\text{comp}}) & \text{Comparison} \\ \mathbb{I}(F_1(f(o_{\text{def}}), y_{\text{def}}) > \alpha) & \text{Defect Adjustment} \\ \mathbb{I}(|o_{\text{score}} - y_{\text{score}}| < \zeta) & \text{Scoring} \end{cases}\)$ Training data: 1,600 slide pairs with controlled perturbations (layout, typography, imagery) annotated by 2–3 human raters.

Key Contributions

  • EvoPresent framework: first multi-agent self-improvement pipeline for academic presentations covering slides, scripts, and narrated video with minimal human intervention.
  • PresAesth: a multi-task GRPO-trained aesthetic model unifying scoring, defect adjustment, and pairwise comparison, replacing VLM-as-judge approaches.
  • EvoPresent Benchmark: 650 papers (13,000 slides) with expert annotations for generation quality; 2,000 slide pairs (6,000 slides) for aesthetic awareness training/eval.
  • Empirical finding that high-quality feedback—not initial agent capability—is the bottleneck for effective self-correction.

Results

  • EvoPresent-claude-4 achieves the best overall presentation quality: PPL 18.57 vs. GPT-4o baseline 24.32 (−24%), ROUGE-L 16.78 vs. 12.59, Aesthetic score 8.05/10 vs. 7.05.
  • EvoPresent-gpt5 highest aesthetic score (8.15/10) and overall fine-grained score (3.85/5).
  • Compared to best multi-agent baseline (Paper2poster-4o, Aesthetic 7.65): EvoPresent-4o gains +0.17 on aesthetic, +0.04 on narrative.
  • PresAesth scoring MAE ~18% lower on average than closed-source VLMs (GPT-4o, GPT-5, Claude-4-Sonnet) on the aesthetic scoring task.
  • Multi-task GRPO outperforms SFT and single-task RL variants in aesthetic awareness generalization.
  • Content differences across models are small; aesthetic/design is the primary differentiator.

Limitations

  • The HTML-based rendering pipeline is not evaluated for compatibility with standard presentation formats (PPTX); the appendix comparison is not reproduced in the main text.
  • PresAesth is trained on slides from top AI conferences (ICLR, NeurIPS, CVPR, etc.)—domain transfer to other disciplines (medicine, law) is untested.
  • VLM-as-judge is still used for fine-grained 8-dimension content evaluation, inheriting its known consistency issues.
  • The default aesthetic threshold \(S_{th}=8.0\) and maximum iteration count \(T\) are fixed hyperparameters whose sensitivity is deferred to the appendix.
  • Human annotation covers 2–3 annotators per slide; inter-annotator agreement statistics are not reported in the main paper.

Relevance to Harnesses / Meta-Harnesses

EvoPresent is a concrete instantiation of a multi-agent orchestration harness: a fixed sequential pipeline of specialized agents (Storyline → Scholar → Design → Checker) with a learned reward model (PresAesth) providing the inner-loop evaluation signal. The key harness design insight is that self-improvement loops require a reliable, task-specific evaluator—replacing generic VLM-as-judge with a fine-tuned RL model is what makes the feedback loop stable enough to converge. The Checker Agent's algorithm (score → revert-or-refine → select-best) is a canonical harness control flow pattern applicable beyond presentations. The paper also empirically validates a principle relevant to any self-correcting harness: initial agent capability does not predict correction ability, so harness designers must treat feedback quality as a first-class design variable.