TL;DR
With research publication volumes skyrocketing, scientific visibility increasingly hinges on effective social media promotion—a process that currently demands substantial manual effort. This paper introduces AutoPR, formalizing academic self-promotion as an automated multimodal generation task, along with PRBench (a benchmark of 512 paper-post pairs) and PRAgent, a specialized three-stage multi-agent pipeline. Real-world social media deployments show PRAgent boosts total watch time by 604% and likes by 438% compared to direct LLM prompting.
What Problem Does This Paper Address?
Academic publishing is experiencing an unprecedented surge, making it nearly impossible for researchers to track all relevant literature through conventional venues alone. As a result, scholars rely heavily on platforms like Twitter/X and RedNote to discover work, and authors must actively promote their papers to secure citations and impact.
However, crafting engaging, accurate, and platform-native promotion requires significant human time and communication expertise. Off-the-shelf Large Language Models (LLMs) struggle when prompted naively: they suffer from severe factual omissions (missing over 40% of core contributions), generate dry, unengaging "announcements" lacking narrative hooks, and use superficial, generic hashtags that fail to reach niche communities.
Methods
The authors formalize AutoPR as a multi-objective optimization problem balancing three axes: Fidelity (factual correctness), Engagement (narrative hooks, clarity, calls-to-action), and Alignment (platform affordances, formatting, hashtags).
To solve this, they propose PRAgent, an end-to-end multi-agent framework operating in three stages:
- Content Extraction & Structuring: A textual agent applies layout-aware parsing and adaptive hierarchical summarization to handle long contexts, while a visual agent uses
DocLayout-YOLOand nearest-neighbor pairing to extract figures, tables, and captions. - Multi-Agent Content Synthesis: Four collaborating agents draft the post:
- Logical Draft Agent: Synthesizes a structured, style-agnostic technical brief (problem, contributions, method, implications).
- Visual Analysis Agent: A multimodal LLM interprets extracted figures and their narrative role.
- Textual Enriching Agent: Rewrites the logical brief into engaging social media prose.
- Visual–Text Interleaved Agent: Dynamically weaves text and selected high-impact images into a cohesive story.
- Platform-Specific Adaptation: An orchestration agent finalizes formatting, tailoring tone, emojis, and hashtags for specific platforms (e.g., Twitter threads vs. visual-heavy RedNote posts).
Datasets
The authors construct PRBench, linking 512 peer-reviewed computer science papers (arXiv 2024–2025 across AI, NLP, and CV) to authentic human-written promotional posts from Twitter/X and RedNote.
- For rapid experimentation, they sample PRBench-Core (128 stratified examples).
- Evaluation combines a weighted human-curated factual checklist with a calibrated LLM-as-a-judge (
Qwen-2.5-VL-72B-Ins, validated against expert annotators with Krippendorff’s $\alpha$ up to 0.93). - In supplementary experiments, they also curate a 24-paper biomedical promotion dataset to test domain generalization.
Key Insights and Ideas
- Inference-Time Scaling Fails for PR: Unlike math or coding, increasing reasoning tokens (Long CoT) does not improve promotional quality. In fact, think token count negatively correlates with final scores ($r = -0.1616$), as over-reasoning introduces "specification drift" and conversational fluff.
- Off-the-Shelf Factuality Metrics Break Down: Standard NLI summarization metrics (like SummaC) perform near chance level ($\approx 0.20$) on AutoPR posts. Because promotion naturally demands analogies, contextualizing framing, and cross-lingual translation (e.g., English papers to Chinese RedNote posts), summarization metrics incorrectly penalize non-literal framing as hallucinations.
- Decomposition Beats Raw Model Scale: While larger parameter scale generally helps, multi-agent task decomposition matters more: PRAgent with mid-sized open models consistently beats direct prompting on top closed-source models like GPT-5.
Key Results
- Benchmark Superiority: Across diverse models (GPT-5, Gemini 2.5, Qwen2.5-VL, InternVL3), PRAgent outperforms direct prompting by 7% to 20%+ on PRBench-Core. On GPT-5, it raised overall quality from 63.97 to 76.62.
- In-the-Wild Field Experiments:
- RedNote (10-day A/B test): PRAgent-generated posts achieved a +604% increase in total watch time, +438% more likes, and a +294% lift in saves/bookmarks compared to standard LLM direct-prompting.
- Twitter/X (7-day test): PRAgent accrued 195 impressions (vs. 56) and generated 170 active engagements (vs. 0 for baseline).
- Human Preference: Human evaluators preferred PRAgent posts over human-authored originals in 64.8% of CS cases (GPT-5 backbone) and 79.2% of biomedical cases.
Limitations and Future Work
- Subjective Visual Evaluation: While LLM judges correlate well with human judgment on text and tagging, aesthetic evaluation of images remains challenging (GPT-4o visual correlation was $<0.1$).
- Lack of Closed-Loop Optimization: PRAgent currently operates open-loop without dynamically adapting to real-time viewer feedback (clicks, likes, comments) via reinforcement learning or bandit algorithms.
- Causality on Citations: Due to the short duration of field experiments, the study demonstrates immediate engagement gains but cannot prove long-term causal increases in academic citations.
- Small Model Gap: When deployed with smaller 7B models, PRAgent still struggles to beat experienced human science communicators (winning only 36.7% of comparisons).
Summary updated