Research map / Multi-agent systems and games / LLM agents
Presenting a Paper is an Art: Self-Improvement Aesthetic Agents for Academic Presentations
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes EvoPresent, a self-improving multi-agent framework that turns papers into narrated presentation videos, guided by PresAesth, a Qwen2.5-VL-7B aesthetic model trained with multi-task GRPO, plus a 650-paper benchmark and 2,000-pair aesthetic dataset.
Problem
Automated academic presentation generation suffers from weak narrative coherence, poor aesthetic quality, and no reliable self-correction, partly because existing evaluators (VLM-as-judge) lack aesthetic awareness.
Approach
A draft-feedback-revision agent loop where a multi-task GRPO-trained aesthetic model (scoring, defect adjustment, pairwise comparison) supplies the feedback signal, plus a benchmark that jointly trains and evaluates aesthetic awareness.
Key results
- PresAesth reaches 87.8% comparison accuracy on the 400-pair test set, versus 77.1% for the best baseline (GPT-4o), and MAE 1.33 for scoring.
- EvoPresent-claude-4 achieves the best overall fine-grained score (3.90 design, 3.90 overall) and ROUGE-L 16.78 among generation methods.
- With PresAesth as checker, EvoPresent-4o raises aesthetic score from 3.2 to over 8.0 within three iterations.
Surprising findings
- Agents' initial performance does not correlate with self-correction ability; stronger models like Gemini-2.5-pro and Claude-4-Sonnet show limited self-correction gains.
- Multi-task SFT achieves higher comparison accuracy than single-task GRPO, while multi-task GRPO gives the best scoring and adjustment results.
Limitations
- Aesthetic training data is limited (about 3,400 instances), and the benchmark relies on about 30 volunteer annotators with inconsistent image sizes handled by resizing.
- Content delivery and video evaluation rely on VLM judges (Qwen-Omni-7B, GPT-4o) rather than humans for most metrics.
- Some dataset and method details are unclear, including the AesthEval/PresAesth naming inconsistency and the perplexity definition that depends on slide images.
Experimental setup
Datasets: EvoPresent Benchmark academic slides SlideAudit
Benchmarks: EvoPresent Benchmark PaperQuiz
Metrics: PPL ROUGE-L Layout Balance Aesthetic score (PresAesth) VLM-judge fine-grained 1-5 scores MAE F1-score Accuracy
Baselines: GPT-4o GPT-5 Claude-4-Sonnet DeepSeek-R1 GPT-4o-Image PPTAgent PresentAgent Paper2Poster VLAA-Thinker-7B Qwen2.5-VL-32B GLM-4.5V
Builds on: GRPO PPTAgent Paper2Poster PresentAgent
Closest papers by meaning
- Paper2Poster: Towards Multimodal Poster Automation from Scientific Papers (NeurIPS 2025 · similarity 0.86)
- P2P: Automated Paper-to-Poster Generation and Fine-Grained Benchmark (ICLR 2026 · similarity 0.85)
- JarvisArt: Liberating Human Artistic Creativity via an Intelligent Photo Retouching Agent (NeurIPS 2025 · similarity 0.85)
- Can Vision–Language Models Assess Graphic Design Aesthetics? A Benchmark, Evaluation, and Dataset Perspective. (ICLR 2026 · similarity 0.84)
- PPT-Eval: A Benchmark for Computer-Use Agents on PowerPoint Tasks (ICML 2026 · similarity 0.84)
- POSTA: A Go-to Framework for Customized Artistic Poster Generation (CVPR 2025 · similarity 0.83)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.