LLMs as Science Journalists: Supporting Early-stage Researchers in Communicating Their Science to the Public

TL;DR

Early-stage researchers face difficulty communicating their technically dense work to the public. The paper introduces a training framework that teaches LLMs to act as science journalists, using data synthesized from papers and press releases to train via supervised fine-tuning and preference learning (SFT and DPO). It demonstrates that journalist-aligned LLMs ask more balanced questions about societal impact, scientific context, and accessibility, outperforming general-purpose baselines in automatic evaluation, and are preferred by PhD students in a user study. The work provides a practical path for hands-on public-science communication training and shares data, models, and interfaces publicly for broader adoption.

Abstract

The scientific community needs tools that help early-stage researchers effectively communicate their findings and innovations to the public. Although existing general-purpose Large Language Models (LLMs) can assist in this endeavor, they are not optimally aligned for it. To address this, we propose a framework for training LLMs to emulate the role of a science journalist that can be used by early-stage researchers to learn how to properly communicate their papers to the general public. We evaluate the usefulness of our trained LLM Journalists in leading conversations with both simulated and human researchers. %compared to the general-purpose ones. Our experiments indicate that LLMs trained using our framework ask more relevant questions that address the societal impact of research, prompting researchers to clarify and elaborate on their findings. In the user study, the majority of participants who interacted with our trained LLM Journalist appreciated it more than interacting with general-purpose LLMs.

Figures (9)

• Figure 1: Example conversation on a paper by asadi2020efficacy, as simulated by Deepseek-R1 model guided by this "https://medicalxpress.com/news/2020-09-surgical-n95-masks-block-particles.html"
• Figure 2: Our training framework: Starting with a corpus of scientific papers and their press releases, we use Deepseek-R1 to synthesize conversations. Second, we use supervised fine-tuning (SFT) to train an LLM to act like a journalist. Third, we synthesize preference data favoring follow-up and societal questions. Finally, the preference data is used to perform preference learning (DPO) over the SFT LLM.
• Figure 3: Aggregation of participants' scoring of their experience interacting with each of the Systems
• Figure 4: Statistics of information overlap between the user-written lay summary and their system interactions, computed using AlignScore (Align.) zha2023alignscore, ROUGE-1 (R1) and ROUGE-L (RL) lin2004rouge.
• Figure 5: Prompts used to generate the preference data to train our LLM Journalist models.