* Equal contribution - Contact: qgchen@ir.hit.edu.cn, zyan@ir.hit.edu.cn, car@ir.hit.edu.cn, lbqin@csu.edu.cn
Formalizes automatic translation of scholarly papers into channel-aware promotion content optimized for fidelity, alignment, and engagement.
512 expertly curated paper-to-post pairs across platforms, complete with weighted factual checklists and human preference judgements.
Three-stage agentic pipeline that boosts watch time by 604% and likes by 438% over strong LLM and rule baselines.
Curious about how AutoPR performs in practice? Visit our Showcase for detailed case studies.
As the volume of peer-reviewed research surges, scholars increasingly rely on social platforms for discovery while authors invest significant effort in promotion to sustain visibility and citations. We introduce Automatic Promotion (AutoPR), a new task that translates research papers into faithful, engaging, and well-timed public-facing content. To enable rigorous study, we release PRBench, a multimodal benchmark linking 512 peer-reviewed articles to high-quality promotional materials and evaluating systems along three axes: Fidelity (accuracy and tone), Engagement (stakeholder targeting and appeal), and Alignment (timing and channel optimization). We further present PRAgent, a modular agentic framework that automates scholarly promotion in three stages: hierarchical content extraction with multimodal preparation; collaborative multi-agent synthesis for polished, publication-ready outputs; and platform-specific adaptation that models norms, tone, and tagging to maximize reach. Against strong LLM pipelines and rule-based tools on PRBench and downstream social metrics, PRAgent delivers substantial gains, including up to a 604\% increase in total watch time, a 438\% increase in likes, and at least a 2.9x rise in overall engagement. Ablations attribute the largest improvements to platform modeling and targeted promotion. Our results establish AutoPR as a tractable, measurable research problem and chart a path toward scalable, trustworthy, and impactful automated scholarly communication.
AutoPR frames scholarly promotion as conditional generation over rich research assets. Each instance starts from a research dossier \(\mathbb{D} = (D_T, D_V, D_S)\) that bundles the full manuscript, figure-caption pairs, and supplementary materials with curated talking points. A dissemination target \((\mathbb{T}_P, \mathbb{T}_A)\) specifies the delivery platform and intended audience persona, grounding tone, cadence, and visual affordances.
\[\hat{P} = \operatorname*{argmax}\limits_{P} \Pr\big(P \mid \mathbb{D}, \mathbb{T}_P, \mathbb{T}_A\big)\]
The generator seeks a Pareto-efficient post \(\hat{P}\) that negotiates competing objectives. AutoPR scores candidate posts with a triad of metrics that reward trustworthy science communication while honoring channel norms.
\[\vec{F}(\hat{P}) = \alpha_1 \mathcal{S}_{\text{Fidelity}}(\hat{P} \mid \mathbb{D}) + \alpha_2 \mathcal{S}_{\text{Align}}(\hat{P} \mid \mathbb{T}_P) + \alpha_3 \mathcal{S}_{\text{Engage}}(\hat{P} \mid \mathbb{T}_A)\]
Balancing these weighted objectives yields a frontier of diverse promotional narratives, ranging from expert-facing summaries to public-friendly explainers, ready for downstream adaptation by PRAgent.
PRBench evaluates automatic promotion systems with expert-curated scores that cover the full lifecycle of science communication.
Every sub-score is assigned by three trained annotators, with disagreements reconciled to build a dependable gold standard for PRBench and its stratified PRBench-Core subset.
PRAgent is a modular agentic workflow spanning three stages:
Direct prompting across the 512-sample PRBench benchmark shows that frontier LLMs still fall short of human-crafted promotion posts. The companion PRBench-Core subset (128 stratified samples) exposes the same failure patterns while enabling rapid iteration.
To score these behaviors at scale we adopt Qwen-2.5-VL-72B-Inst. as the automatic judge. Its decisions correlate with human ratings up to 0.98 on factual accuracy and 0.75 on authorship fidelity, giving us reliable signal across PRBench and PRBench-Core.
LLM failures on PRBench concentrate on missing precise facts, over-using generic engagement hooks, and diverging from human hashtag choices.
PRAgent orchestrates content extraction, multi-agent synthesis, and platform adaptation to remediate the above shortcomings. We evaluate 12 competitive LLM backbones on PRBench-Core for fast comparison and confirm the same ordering on the full 512-sample PRBench.
Top: PRAgent lifts fidelity, engagement, and alignment metrics over direct prompting for every backbone on PRBench-Core. Bottom: The same ordering holds on the full 512-sample PRBench benchmark.
To validate practicality we ran a 10-day RedNote study with paired accounts posting identical daily papers. PRAgent powered one account, while the baseline relied on direct prompting with the same GPT-5 backbone.
Real-world RedNote deployment: PRAgent drives higher interactions per paper and sustained audience growth compared to direct prompting.
If you find AutoPR helpful, please cite our work.
@article{chen2025autopr,
title={AutoPR: Let's Automate Your Academic Promotion!},
author={Chen, Qiguang and Yan, Zheng and Yang, Mingda and Qin, Libo and Yuan, Yixin and Li, Hanjing and Liu, Jinhao and Ji, Yiyan and Peng, Dengyun and Guan, Jiannan and Hu, Mengkang and Du, Yantao and Che, Wanxiang},
journal={Manuscript},
year={2025},
note={\url{https://autopr.github.io}}
}
Please create Github issues here or email Qiguang Chen or Zheng Yan , or open up an issue on Github if you have any questions or suggestions.
This website is adapted from Nerfies, LLaVA-RLHF, and M3CoT, licensed under a Creative Commons Attribution-ShareAlike 4.0 International License.
Usage and License Notices: The data, code and checkpoints are intended and licensed for research use only. Please ensure compliance with the corresponding licenses when using PRBench, PRAgent, or other AutoPR assets.