AutoPR: Let's Automate Your Academic Promotion!

AI summary

Overview

Research area: Natural Language Processing / AI for Science (AI4Research), specifically automated scholarly communication and multi-agent LLM systems.

Technical level: Intermediate. The paper defines a formal multi-objective task and an evaluation benchmark, but its method is presented at the system-and-pipeline level rather than as new model training.

Scope: The authors formalize "Automatic Promotion" (AutoPR) — automatically turning a peer-reviewed paper into engaging, platform-tailored promotional content — and release both a benchmark (PRBench) and a multi-agent system (PRAgent) for it.

What This Paper Is About

The number of accepted conference papers has grown sharply, making it hard for readers to keep up and harder for authors to get noticed. Authors therefore spend substantial manual time and cost promoting their work on social platforms, yet no systematic task, dataset, or pipeline existed for automating that promotion. This paper defines AutoPR as a research task, builds PRBench to measure it, and proposes PRAgent to perform it end-to-end.

Key Contributions

1. The AutoPR task. The authors formalize automatic academic promotion as translating peer-reviewed research into tailored promotional materials, with specified inputs (manuscripts, figures, key findings) and outputs (press-release-style or social media posts, visual abstracts), evaluated by Fidelity, Alignment, and Engagement metrics.
2. PRBench. A publicly released benchmark of 512 paired multimodal samples linking peer-reviewed papers to human-authored promotional posts across three AI-related fields, constructed from arXiv papers submitted between June 2024 and June 2025 and posts from Twitter (X) and RedNote. A 128-sample stratified subset, PRBench-Core, enables faster, cheaper evaluation.
3. PRAgent. A three-stage multi-agent framework combining Content Extraction, Multi-Agent Content Synthesis, and Platform-Specific Adaptation. The authors report it outperforms direct LLM prompting baselines across nearly all evaluated LLMs, by at least 7.15% on nearly all models and metrics.
4. Real-world validation. A 10-day in-the-wild study on RedNote with two new accounts (PRAgent-generated content versus a direct-prompt baseline), in which PRAgent posts achieved at least 294% improvement in interaction metrics, with total watch time up 604% and profile visitors up 575%.

Main Findings

• Current LLMs struggle on academic promotion. Across evaluated models, average scores on PRBench-Core range from 31.05 to 70.56, including for GPT-5.
• Fidelity is a bottleneck. Even Qwen-2.5-VL-32B-Ins, one of the stronger models, scores only 59.87 on Factual Score, missing over 40% of key facts. In 512 outputs from that model, over 92% of errors fall into Numerical/Method/Terminology categories.
• Engagement is formulaic rather than genuine. Analysis of hook strategies shows 42% of AI-generated posts lack any engagement device, and models tend to produce announcements rather than narrative-driven writing.
• Platform alignment is superficial. The average Jaccard similarity between generated and human hashtags was only 0.03, indicating reliance on generic high-frequency tags rather than niche keywords.
• PRAgent improves results substantially. On GPT-5-mini, PRAgent improvements surpass 20%; gains over the direct-prompt baseline are at least 7.15% across nearly all models and metrics. PRAgent is reported to surpass human authors in preference studies.
• Real-world gains. In the 10-day RedNote study, total watch time increased by 604%, likes by 438%, profile visitors by 575%, and overall engagement by at least 2.9x relative to the baseline. The paper summarizes this in the contributions as up to a 6x increase in total watch time and a 4x increase in likes.
• Ablations. Platform modeling and targeted promotion contribute the most to PRAgent's gains; all three stages of the framework are reported as essential.

Methodology in Plain English

The work has three parts.

Task and benchmark. An input paper is represented as its text, a set of figure/table images with captions, and supplementary materials. The target is described by a platform (for example Twitter or RedNote) and an intended audience. The system must produce a post that maximizes a weighted combination of three scores: Fidelity (factual accuracy and completeness relative to the paper), Alignment (fit with platform norms of style, tone, format, hashtags), and Engagement (potential to attract and resonate with the target audience).

Benchmark construction. Papers were collected from arXiv (June 2024 to June 2025, in Computation & Language, Machine Learning, and Artificial Intelligence), and matching promotional posts were retrieved from Twitter (X) and RedNote. Posts were filtered for AI-generated content, then 512 parallel paper–post pairs were sampled and manually verified. Human experts built weighted factual checklists (with Gemini 2.5 Pro drafting an initial checklist that experts corrected) and rated posts on 0–5 scales; each post was independently assessed by a panel of at least three annotators, with small discrepancies averaged and larger ones resolved by consensus.

PRAgent pipeline.

• Stage 1, Content Extraction: text is converted to HTML with PyMuPDF and hierarchically summarized (single-pass for short text, recursive section-by-section summarization for long text). In parallel, PDF pages are rendered to 250 DPI PNG images, layout is analyzed with DocLayout-YOLO to detect figures, tables, and captions, and a nearest-neighbor algorithm pairs visuals with their captions.
• Stage 2, Multi-Agent Content Synthesis: four agents cooperate. A Logical Draft Agent turns the summary into a structured, style-agnostic draft covering the research question, contributions, method, and results. A Visual Analysis Agent uses a multimodal model to interpret each figure. A Textual Enriching Agent rewrites the draft into platform-native text with hooks, calls-to-action, and hashtags. A Visual-Text-Interleaved Combination Agent decides where images will have the most impact and inserts placeholders.
• Stage 3, Platform-Specific Adaptation: an Orchestration Agent rewrites the draft for the target platform's tone, formatting, emojis, and hashtags, then replaces placeholders with Markdown image tags and bundles the output with its image assets.

Evaluation uses Qwen-2.5-VL-72B-Ins as the LLM judge, selected because it showed the strongest and most consistent correlation with human annotations.

Why This Matters

Impact on research. The paper opens a previously underexplored stage of AI4Research: research promotion and dissemination. Prior work on automated posters and science journalism exists, but a standardized benchmark and end-to-end pipeline did not. PRBench provides a measurable target for future work, and the authors' analysis shows that generic summarization is insufficient for effective academic promotion.

Real-world applications.

• Individual researchers promoting papers on platforms like Twitter (X) and RedNote without manual post-writing.
• University and lab communications teams producing visual abstracts and social posts at scale.
• Publishers and journals disseminating published articles to broader, non-specialist audiences.
• Researchers seeking to reach specific stakeholder groups such as journalists, policymakers, or industry practitioners.

Industry relevance. Automated, platform-aware content generation from technical source material is directly relevant to scientific publishing, research-marketing tools, altmetrics and analytics services, and any organization that needs to translate dense technical documents into audience-specific communication.

Future Directions

• Improving factual fidelity, since even the best-performing models miss a large share of key facts and most errors are numerical, methodological, or terminological.
• Making engagement more genuinely human-like, given that a large fraction of generated posts contain no engagement device and read like announcements.
• Deeper platform modeling, since hashtag overlap with human posts is very low and current models rely on generic high-frequency tags.
• Extending evaluation and validation beyond the 10-day RedNote study and the two social platforms used, and beyond the three AI-related paper fields represented in PRBench.

Target Audience

This paper is most useful to researchers working on multi-agent LLM systems, AI for science, and text generation evaluation; to developers building publication-support or research-marketing tools; and to science communicators, lab communications staff, and individual academics who want a systematic, measurable approach to promoting technical work on social platforms.

Authors’ abstract

As the volume of peer-reviewed research surges, scholars increasingly rely on social platforms for discovery, while authors invest considerable effort in promoting their work to ensure visibility and citations. To streamline this process and reduce the reliance on human effort, we introduce Automatic Promotion (AutoPR), a novel task that transforms research papers into accurate, engaging, and timely public content. To enable rigorous evaluation, we release PRBench, a multimodal benchmark that links 512 peer-reviewed articles to high-quality promotional posts, assessing systems along three axes: Fidelity (accuracy and tone), Engagement (audience targeting and appeal), and Alignment (timing and channel optimization). We also introduce PRAgent, a multi-agent framework that automates AutoPR in three stages: content extraction with multimodal preparation, collaborative synthesis for polished outputs, and platform-specific adaptation to optimize norms, tone, and tagging for maximum reach. When compared to direct LLM pipelines on PRBench, PRAgent demonstrates substantial improvements, including a 604% increase in total watch time, a 438% rise in likes, and at least a 2.9x boost in overall engagement. Ablation studies show that platform modeling and targeted promotion contribute the most to these gains. Our results position AutoPR as a tractable, measurable research problem and provide a roadmap for scalable, impactful automated scholarly communication.