PRAgent: Automated Academic Promotion

• PRAgent is a multi-agent framework that automates the promotion of academic research by converting full papers into concise, engaging, and platform-optimized posts using advanced extraction and synthesis.
• The framework utilizes specialized agents for content extraction, multimodal integration, and targeted adaptation, resulting in significant improvements in engagement metrics such as watch time and likes.
• Empirical evaluation on the PRBench benchmark and ablation studies confirm PRAgent’s scalability and efficacy in enhancing research visibility and impact.

PRAgent is a multi-agent framework designed to automate the transformation of academic research papers into engaging, accurate, and platform-optimized promotional posts, as described in the AutoPR system. It targets the increasing demand for scalable scholarly communication on social platforms by decomposing the promotion process into specialized agent stages, combining content extraction, multimodal synthesis, and nuanced adaptation to target audience and platform norms.

1. System Overview and Structural Design

PRAgent operationalizes academic promotion in three sequential stages, each handled by a distinct set of agents:

1. Content Extraction and Structuring: The system ingests the full research document, including textual and visual sources (figures, tables, supplemental items). The Textual Content Extraction Agent employs hierarchical summarization, parsing the document into an intermediate HTML structure and summarizing by section so that even long documents are compressed to fit within LLM context limitations. The Visual Content Preparation Agent extracts visuals via image conversion and layout segmentation (e.g., DocLayout-YOLO), pairing each with its associated caption or figure description.
2. Multi-Agent Collaborative Synthesis:

Specialist agents operate in parallel and in sequence: - The Logical Draft Agent generates a fact-centric narrative, filling a fixed schema (research question, contribution, method, import). - The Visual Analysis Agent uses a multimodal LLM to analyze each extracted figure, producing a description of content and scientific value. - A Visual-Text-Interleaved Combination Agent merges the logical draft and visual commentary into one cohesive, visually grounded narrative.

1. Platform-Specific Adaptation and Orchestration: The Orchestration Agent tailors the final product to target platform conventions (e.g., Twitter, RedNote), optimizing for tone, formatting, use of hashtags, emoji, and attributions. This last stage ensures outputs are not only scientifically faithful but are also aligned with audience targeting, platform culture, and maximum potential reach.

This modular, decomposition-based architecture allows each agent to focus on a well-defined subtask, facilitating scalability and system extensibility for future addition of platform types or adaptation strategies.

2. Metrics and Quantitative Outcomes

Evaluation on PRBench, a dedicated multimodal benchmark, demonstrates that PRAgent substantially outperforms direct LLM pipelines:

Metric	PRAgent Performance Improvement
Total Watch Time	+604% compared to direct prompting
Likes	+438%
Overall Engagement	≥2.9× (including broader audience, click-through, social interaction)

Performance gains are observed across fidelity, engagement, and alignment axes in both core and extended PRBench scenarios. For example, fidelity is measured by an LLM-based factual checklist score:

$$\mathcal{S}_\text{Checklist}(P \mid \mathbb{D}) = \frac{\sum_{i=1}^{n} w_i \cdot v(P \mid c_i, \mathbb{D})}{\sum_{i=1}^{n} w_i}$$

where $w_i$ is the weight, and $v$ is an LLM-driven verification function for the i-th factual requirement.

Empirical ablation experiments confirm the necessity of each agent stage: omitting content extraction, collaborative synthesis, or adaptation and orchestration causes significant drops in fidelity, alignment, and engagement scores.

3. Formalization of the Promotional Generation Process

The end-to-end transformation is modeled as a multi-objective optimization, in which the system searches for the promotional post $P$ that maximizes joint objectives:

$$\hat{P} = \operatorname*{argmax}_{P}\, \Pr(P \mid \mathbb{D}, \mathbb{T}_P, \mathbb{T}_A)$$

with the overall scoring function

$$\alpha_1\, \mathcal{S}_\text{Fidelity}(P \mid \mathbb{D}) + \alpha_2\, \mathcal{S}_\text{Align}(P \mid \mathbb{T}_P) + \alpha_3\, \mathcal{S}_\text{Engage}(P \mid \mathbb{T}_A)$$

where $\alpha_1, \alpha_2, \alpha_3$ are weights for fidelity, platform alignment, and audience engagement, $\mathbb{D}$ is the document, $\mathbb{T}_P$ the platform template, and $\mathbb{T}_A$ the audience template.

4. The PRBench Evaluation Benchmark

PRBench is a highly curated resource that underpins measurable progress in AutoPR research:

• Composition: 512 academic articles (full text plus visuals) paired with manually selected, platform-optimized promotional posts.
• Axes of Evaluation:
  • Fidelity (completeness, accuracy, tone, correct claims)
  • Engagement (audience targeting, hook/CTA, appeal)
  • Alignment (timing, platform norms, formatting, hashtag/mention use, and visual-text integration)
• Methodology:

The benchmark employs both quantitative (e.g., factual checklist, engagement score) and qualitative (preference, relevance, platform fit) instruments for comprehensive appraisal.

5. Ablation and Module Contribution Analysis

Ablation studies isolate the contribution of each subsystem to overall performance:

Agent Stage Omitted	Major Observed Effects
Content Extraction (Stage 1)	Significant fidelity loss (e.g., score drops from 70.76 to 66.38)
Multi-Agent Synthesis (Stage 2)	Alignment and narrative coherence degradation
Platform Adaptation (Stage 3)	Severe misalignment with promotional norms; lowest overall scores

Findings attribute most of the gain to the platform adaptation and enhanced content extraction components.

6. Practical Implications and Outlook

PRAgent demonstrates that decomposing promotion into specialized stages—with each agent optimized for a specific function and using both textual and visual cues—yields scalable, measurable, and high-engagement academic promotion.

Implications include:

• Scalability: Automation reduces the manual overhead of research dissemination, particularly valuable given the growth rate of peer-reviewed publication volume.
• Research Impact: By increasing engagement and reach (as evidenced by strong metrics), PRAgent has the potential to amplify the influence of new scientific discoveries.
• Measurability: The introduction of PRBench allows rigorous, repeatable comparisons and will support the development of more sophisticated agents.

Prospective research directions involve more adaptive visual selection, dynamic inter-agent communication, rapid fine-tuning for emerging platforms, and the refinement of engagement and aesthetic metrics in conjunction with human evaluation.

7. Summary Table: PRAgent Stages and Functions

Stage	Key Agent Functionality	Technical Methods
Content Extraction & Structuring	Hierarchical summarization, visual extraction	LLM summarization, DocLayout-YOLO segmentation
Collaborative Content Synthesis	Logical narration, visual-text synthesis	Multimodal LLMs, schema-based drafting
Platform-Specific Adaptation	Platform norm modeling, tone/formatting	Dedicated Orchestration Agent

The cumulative design advances the tractability and impact of automated scholarly communication, substantiated by strong empirical results on PRBench and methodical ablation analysis (Chen et al., 10 Oct 2025).