RevTogether: Supporting Science Story Revision with Multiple AI Agents Conference: ; ; CCS: Human-centered computing Interactive systems and tools
Yu Zhang, Kexue Fu, Zhicong Lu · 2025
Read RevTogether: Supporting Science Story Revision with Multiple AI Agents Conference: ; ; CCS: Human-centered computing Interactive systems and tools by Yu Zhang, Kexue Fu, Zhicong Lu (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2503-01608.txt.
Introduction
Yu Zhang, Kexue Fu, and Zhicong Lu build a multi-agent writing assistant for science stories in a 2025 paper. Science stories mix narrative charm with accurate knowledge, but crafting them takes skill. Writers must reshape complex concepts for readers with very different science backgrounds.
RevTogether assigns several AI agents, built on GPT-4o, to play human-like roles in revision. The agents simulate readers and editors so a writer can see how revisions land before publishing. Supporting revision specifically matters because first drafts of popular science rarely survive contact with real audiences unchanged.
Catalog record
| CREATOR | Yu Zhang, Kexue Fu, Zhicong Lu |
|---|---|
| YEAR | 2025 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 28 KB |
| VERSION | 1.0 |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2503-01608
Browse the full library · About the Tower of Babel
Full text
The opening of the work, from the library's hosted text:
RevTogether: Supporting Science Story Revision with Multiple AI Agents Conference: ; ; CCS: Human-centered computing Interactive systems and tools Yu Zhang, Kexue Fu, Zhicong Lu Abstract . As a popular form of science communication, science stories attract readers because they combine engaging narratives with comprehensible scientific knowledge. However, crafting such stories requires substantial skill and effort, as writers must navigate complex scientific concepts and transform them into coherent and accessible narratives tailored to audiences with varying levels of scientific literacy. To address the challenge, we propose RevTogether, a multi-agent system (MAS) designed to support revision of science stories with human-like AI agents (using GPT-4o). RevTogether allows AI agents to simulate affects in addition to providing comments and writing suggestions, while offering varying degrees of user agency. Our preliminary user study with non-expert writers (N=3) highlighted the need for transparency in AI agents’ decision-making processes to support learning and suggested that emotional interactions could enhance human-AI collaboration in science storytelling. Figure 1. Overview of the RevTogether System. (A) The writer can ask the commentator agent for feedback on a randomly selected segment of the texts. (B) The commentator agents present human-like avatars with emotional appearances matching the sentiment of comments. (C) The writer can review and respond to each comment. (D) Once the writer agrees with a comment, a writing assistant agent will provide writing technique suggestions for addressing the problems mentioned in comments. (G) The writer can further choose a technique to see (E) which parts of the texts are suggested to apply the technique and (F) review detailed revision suggestions provided by AI agents. Overview of the \system{} System. (A) The writer can ask the commentator agent for feedback on a randomly selected segment of the texts. (B) The commentator agents present human-like avatars with emotional appearances matching the sentiment of comments. (C) The writer can review and respond to each comment. (D) Once the writer agrees with a comment, a writing assistant agent will provide writing technique suggestions for addressing the problems mentioned in comments. (G) The writer can further choose a technique to see (E) which parts of the texts are suggested to apply the technique and (F) review detailed revision suggestions provided by AI agents. 1. Introduction Science stories have long captivated broad audiences with varying levels of scientific literacy by blending engagement, entertainment, and comprehensible scientific knowledge ( Davies et al. 2019 ; Joubert et al. 2019 ; Zhang et al. 2023 ) . However, crafting quality stories from complex scientific facts highly relies on writers’ expertise and mastery of advanced writing strategies ( ElShafie 2018 ; August et al. 2020 ; He et al. 2024 ) . For example, even scientists, who are usually the primary source of new scientific discoveries, also struggle to write engaging and accessible narratives to share their research insights with the general public ( Tuttle et al. 2023 ; Côté and Darling 2018 ; Smith et al. 2018 ) . Fortunately, advanced computing and artificial intelligence (AI) technologies could bring opportunities to lower the barriers to science story creation, empowering a wider range of individuals to participate in science communication. With the rapid development of Large Language Models (LLMs) ( Brown et al. 2020 ) and Generative Artificial Intelligence (GenAI) ( Muller et al. 2022 ) , researchers in multiple disciplines including Human-computer Interaction (HCI) have noticed the potentials of applying AI to support various tasks in science communication ( Alvarez et al. 2024 ; Biyela et al. 2024 ; Nishal et al. 2024 ; Williams et al. 2022 ) , including writing for audiences from the general public ( Jiang et al. 2024 ; Gero et al. 2023 ; Schäfer et al. 2024 ; Tatalovic 2018 ) . Previous work has explored approaches to scaffolding the creation of science news and social media content using AI ( Kim et al. 2023 ; Gero et al. 2022 ; Laban et al. 2024 ) . For example, LLMs are used to apply particular writing strategies (e.g., “hook” in Long et al. ( Long et al. 2023 ) and “metaphor” in Kim et al. ( Kim et al. 2023 ) ), provide creative inspirations ( Gero et al. 2022 ) , and retrieve additional information from external sources ( Laban et al. 2024 ) . However, these approaches neglect feedback from other perspectives, focusing solely on fulfilling the writer’s demands. Feedback from different perspectives is essential in writing because it helps to build an iterative revision process for writers to continuously improve the outcomes ( Benharrak et al. 2024 ; Du et al. 2022 ; Nelson and Schunn 2009 ) . Lacking feedback prevents writers from making changes with clear directional guidance ( Long et al. 2023 ; Wigglesworth and Storch 2012 ) , especially for science stories that can have a wide range of readers from scientists to the general public ( ElShafie 2018 ; Burns et al. 2003 ) . Inspired by Benharrak et al.’s previous research that highlights the potential of social factors (e.g., relation, valence) in designing AI personas for writing feedback ( Benharrak et al. 2024 ) , we propose RevTogether, a multi-agent system (MAS) that supports writers in revising science stories together with two human-like commentator AI agents and a writing assistant agent. Specifically, we designed the personas of the two commentator agents as a “mad scientist” and a “curious girl”, representing the two sides of the “dialogue model” in science communication—scientists and ordinary people ( Trench 2008 ; Reincke et al. 2020 ) —for providing feedback in diverse perspectives. To enhance emotional expression, we introduced character avatars with dynamic emotional responses for the commentator agents in RevTogether, providing non-verbal feedback to engage the writers. Through a preliminary user study with three non-expert writers, we highlighted the importance of the transparency of AI agent’s suggestion generation process because it could be helpful for writers who want to learn from AI. This finding aligned with the previous work on human-AI alignment in writing tasks ( Wang et al. 2025 ; Gero et al. 2023 ; Goyal et al. 2024 ; Chakrabarty et al. 2024 ) , revealing learning as a new demand in human-AI collaborative writing. Besides, we also found that emotional reactions simulated by AI agents could serve as effective feedback to stimulate writers’ thoughts in science storytelling. Our findings could inspire future research on AI agent design for broad tasks in content creation in the context of science communication. 2. System Design and Implementation RevTogether aims to support science story revision with on-demand feedback provided by two commentator agents and a writing assistant agent. To begin with, we chose the Human-AI teaming (HAT) analysis framework ( Newendorp et al. 2024 ; Sepich et al. 2021 ) as design guidelines to operationalize our idea into a feasible prototype system. According to the HAT framework, RevTogether is a multi-agent-single-human system in which the two commentator agents and the writing assistant agent collaborate and interact with a single human writer. With different roles played and tasks performed by the agents, RevTogether offers multiple levels of user agency. The overall layout of RevTogether’s user interface (UI) is similar to common document editing tools (e.g., Microsoft Word and Google