PaperlayerView exact revision on arXiv

Source-linked AI summary

Teachers' perspective on AI-based Multi-Agent Simulation Design to Combat School Bullying

Jiaju Lin, Ellen Wenting Zou, Feiwen Xiao, Huanying Song

arXiv:2609.35776v1cs.HCcs.CY

TL;DR

Existing programs lack long-term effects due to insufficient customization and stakeholder involvement. The paper co-designed a multi-agent simulation system with teachers, using LLMs to render bullying situations involving multiple participants. Teachers provided positive feedback and some feasible future directions.

Abstract

from arXiv · show

Bullying in schools profoundly affects the mental and physical health of teenagers. Although existing in-person and digital interventions provide some benefits, they often fall short in addressing the complex social dynamics of bullying. In this study, we collaborated with K-12 teachers to co-design a multi-agent anti-bullying system powered by large language models (LLMs). This system simulates authentic scenarios, enabling students to develop anti-bully skills. The research identifies key design parameters for an LLM-driven multi-agent simulation system, offering valuable insights for creating more effective and scalable anti-bullying tools that could significantly reduce bullying in schools} \keywords{anti-bullying interventions, multi-agent system, generative AI, co-design, bystander presence

Keywords: anti-bullying interventions, multi-agent system, generative AI, co-design,

The paper proposes a teacher-informed, LLM-driven multi-agent system to simulate bullying scenarios and support bystander intervention. It frames the system against limits in existing programs’ ability to represent complex social dynamics and customize for complex cases.

Bullying is a pervasive issue significantly impacting the mental health and well-being

initiatives. In-person programs have been refined over decades, showing positive effects

Despite these initiatives, challenges persist. Some in-person anti-bullying programs

  • Existing digital programs and serious games often fail to simulate bullying’s complex multi-participant social dynamics and lack customization for complex cases.The paper also notes that some in-person programs may have limited long-term effects due to insufficient customization and stakeholder involvement.Source: Despite these initiatives, challenges… · p. 3

To address these gaps, this study proposes a multi-agent anti-bullying system design

  • The proposed system uses teachers’ knowledge and experiences with LLMs to simulate complex social dynamics and varied bullying-story progressions, with personalized feedback.Its goal is to identify key design parameters for an LLM-driven multi-agent simulation system.Source: To address these gaps, this study… · p. 4

LLM-driven multi-agent bystander intervention system to combat bullying

  • The study identifies design parameters and guidelines for educators and developers creating engaging applications to shape teens’ prosocial behavior against bullying.The authors present this as a theoretical and practical contribution to anti-bullying education.Source: LLM-driven multi-agent bystander… · p. 4

potentially have a more lasting impact on reducing bullying in schools. 1

The paper positions bystander intervention as important because bystanders’ reactions can affect bullying outcomes. It aims to help teens practice intervention strategies in simulation, addressing limited opportunities for practice in existing programs.

Bystanders, witnesses to bullying incidents, can significantly influence the outcomes

  • In discussing factors that inform potential strategies, Forsberg et al. (2016) identified five factors influencing bystander decisions: informed awareness, behavioral seriousness, personal feelings, bystander expectations, and sense of responsibility.Thornberg & Jungert (2013) identified basic moral sensitivity, moral disengagement, and defender self-efficacy as factors affecting bystander behaviors.Source: Bystanders, witnesses to bullying… · p. 5
  • The study aims to improve bystanders’ self-efficacy by letting teens test intervention strategies in simulation, with the intended benefit of greater awareness and empathy toward victims.The cited background links individual defender self-efficacy with defending behavior (Thornberg et al., 2020).Source: Bystanders, witnesses to bullying… · p. 5

role-playing scenarios to increase empathy towards victims of cyberbullying. Social Media

  • Existing programs were reported effective, but opportunities for bystanders to practice intervention skills are limited.The study responds by constructing a simulated environment with responsive AI characters to help students practice and reflect on intervention strategies.Source: role-playing scenarios to increase… · pp. 5–6

AI-powered systems to detect, prevent, and provide intervention strategies for bullying (Arif,

Chatbots offer immersive conversational learning experiences for anti-bullying

The paper argues that chatbots do not represent the nuanced, changing interactions among multiple participants in bullying scenarios. It recruited teachers with bullying experience and used an initial co-design stage to shape authentic scenarios and intervention mechanisms.

Chatbots offer immersive conversational learning experiences for anti-bullying

  • Prior conversational anti-bullying programs were described as interactive, but with limitations in emotion detection, adapting language for children, and predictability (Lafrance St-Martin & Villeneuve, 2024).A cited example combined a chatbot and serious game to teach bullying and bystander intervention (Sanoubari et al., 2022).Source: Chatbots offer immersive… · p. 6
  • Existing chatbots lack nuanced interactions among multiple participants and real-time changes in their emotional states and overall social dynamics.The study aims to address this gap with an LLM-based multi-agent system informed by educators’ pedagogical insights.Source: Chatbots offer immersive… · pp. 6–7

3.1 Participants and Research Design

  • Fifteen K-12 teachers with experience dealing with school bullying were recruited; thirteen completed the co-design stages, and most taught in mainland China and Hong Kong.Two participants were from the USA and one from Iran.Source: §3.1 · p. 7

1st Co-design stage: understanding the social dynamics of school bullying

  • In the first 60-minute co-design stage, teachers discussed bullying cases, peer-group status and personalities, bystander roles, and effective interventions for de-escalation.The interviews were intended to inform authentic bullying scenarios for the intervention.Source: 1st Co-design stage: understanding the… · p. 7

2nd Co-design stage: explorating a prototype to surface design guidelines

The second co-design stage tested and iteratively refined a teacher-informed prototype. Teachers assessed its design and perceived usefulness and engagement, while interview transcripts were coded to identify bullying dynamics and intervention-design preferences.

2nd Co-design stage: explorating a prototype to surface design guidelines

Our team designed the simulation system using state-of-the-art LLMs, which provide

  • The simulation uses AI agents as the bully, reinforcer, and victim, while the human user takes the bystander role; each agent responds to input from the other roles.Teachers’ scenario elements included characters’ personality traits and social status, student conflicts, and representative bullying comments.Source: Our team designed the simulation… · p. 9
  • In the second design stage, an iterative approach used feedback from each participant batch to refine the system for subsequent groups; participants reported perceived usefulness and engagement based on two scales..839 was reported for perceived engagement.Source: Our team designed the simulation… · p. 9, Cronbach's alpha was calculated and… · p. 10

3.3 Data analysis

  • Three researchers independently open-coded the interview transcripts, then used axial coding to group related codes into overarching themes.They discussed discrepancies until reaching 100% agreement.Source: §3.3 · pp. 10–11
  • The interview analysis yielded insights into bullying social dynamics intended to guide bystander-intervention design.The analysis also identified common bullying patterns and how those patterns informed design goals.Source: §3.3 · p. 11

Profiles of Roles in Bullying Scenarios

Teachers describe bullying as shaped by social status and uneven peer relationships: bullies may be popular, while victims often hold marginal positions. Bystanders—especially peers with relevant ties or leadership roles—can respond when teachers are absent.

4.1 Unpacking social dynamics of bullying among teens

  • Bullies often hold high peer status, while victims commonly occupy weak or marginal positions in social groups.Teachers linked bullies’ popularity to academic achievement, appearance, or class roles; victims’ marginal status was associated with having fewer friends.Source: §4.1 · p. 12
  • Bullies’ pursuit of social status can make behavioral change difficult, as some target perceived competitors and resist outside guidance.Teachers described long-standing reinforcement of bullies’ behavior and noted that some students may struggle to accept a teacher’s perspective.Source: §4.1 · p. 12

discrimination from peers. Many teachers reported that victims suffered from mental health

Bystander intervention is critical due to its timeliness, as teachers aren't always

  • Bystanders can respond promptly when teachers are absent; teachers highlighted empathy, class leaders, bullies’ role models, and victims’ friends as potential interveners.Suggested responses include condemning bullying, helping bullies recognize harm, and supporting victims.Source: §4.1 · p. 12, Bystander intervention is critical due… · p. 13

Teachers suggested we develop the intervention into an educational game to improve

Teachers favored an engaging role-playing game where students can make choices, receive feedback, and experience different roles and outcomes. The proposed design follows experiential learning through conceptual learning, experimentation, and reflection.

Teachers suggested we develop the intervention into an educational game to improve

  • Teachers envisioned an engaging role-playing game with decision-making, timely feedback, and immersive interaction to help students explore bullying responses.They described seeing how interactions across roles lead to different outcomes as a useful feature.Source: Teachers suggested we develop the… · p. 14
  • Experiential learning structures the intervention as a cycle of concrete experience, reflection, conceptualization, and active experimentation.The authors selected this framework to connect learning with real-world experience.Source: Teachers suggested we develop the… · p. 14

In concrete, our intervention should contain three components, covering

Summative analysis to facilitate reflection, completing the loop of experiential

The prototype combines conceptual learning, simulated experimentation, and guided reflection, drawing on teacher feedback to shape its design. Teachers saw educational value in practice and empathy-building, while also raising concerns about student motivation for optional use.

4.3 Translating human insights into multi-agent intervention design

  • The prototype organizes experiential learning into conceptualization, experimentation, and reflection, using AI-character reactions and emotional states as feedback during simulation.Learners can review qualitative and quantitative analyses and use rollback to retry different intervention approaches.Source: §4.3 · p. 16

Based on teachers’ comments, we adopted the core principles of experiential learning

find themselves at a loss for words when they want to speak.”

  • Teachers cautioned that students may not use the system proactively after school, particularly if they have not encountered bullying.They suggested integration into social-emotional learning or semi-compulsory programming, alongside rewards or more attractive customizable scenarios.Source: find themselves at a loss for words… · p. 17
  • To sustain interest, teachers proposed customizable characters and scenarios, familiar story material, and engaging features such as rewards.Source: find themselves at a loss for words… · pp. 17–18

literature to make it more engaging and interesting for the students.”

Teachers generally found the scenario recognizable but wanted more contextual detail. Feedback also shaped real-time emotional cues, reflective analysis, and message pacing, exposing a trade-off between simulation authenticity and cognitive load.

4.3.2 Design based on Teachers’ Experience Guarantees the Authenticity of the Story

  • Teachers found the school scenario plausible but requested plot details such as group size, discussion type, and scoring to support engagement.One teacher said conflicts can arise over small matters, while another warned that missing background detail might lower engagement.Source: §4.3.2 · p. 18

upstanding behaviors

The analysis module is designed to embody the reflective observation in experiential

  • The analysis module added qualitative feedback and alternative expressions after teachers said constructiveness and aggressiveness scores alone were not educationally sufficient.Teachers praised the final analysis system’s detailed suggestions and its support for reflection.Source: The analysis module is designed to… · p. 19

For bullying types, Our collected information reflected this trend, many participants

  • Teachers noted that verbal and social bullying are common, while frequent agent messages can dominate users’ cognitive load.After participants reported overwhelming message flow, the team muted proactive direct messages and slowed AI-agent messages; later participants seldom complained about authenticity.Source: For bullying types, Our collected… · pp. 19–20
  • The design challenge was balancing authentic proactive communication against message overload, with message frequency identified as a major source of cognitive load.Source: For bullying types, Our collected… · p. 19

Change of Mean Perceived Usefulness and Mean Perceived Engagement across all

Perceived usefulness and engagement stayed above “somewhat agree” and generally rose across versions. Teachers linked engagement to interactivity, while identifying abstract instruction and inauthentic AI language as areas for improvement.

4.3.6 Overall perceived engagement and usefulness

  • Both perceived usefulness and engagement remained above 3 (“somewhat agree”) and generally increased as the intervention versions changed.The average engagement score was still described as having room for improvement.Source: §4.3.6 · p. 20, Although the average perceived… · p. 20

Teacher participants indicated that interactivity is the key element that determines

Teachers critiqued our conceptualization part for focusing solely on knowledge

  • Teachers criticized the conceptualization stage for emphasizing knowledge representation and testing while offering too little user involvement.One teacher found the phrases abstract and said the section did not help them think through how to solve the first step.Source: Teachers critiqued our… · p. 21
  • Teachers proposed visual scenes and animations to make conceptual content more engaging, and suggested carrying its intervention concepts into later choices or prompts.The conceptualization stage was intended to help students experience and reflect on abstract bullying-related concepts.Source: Teachers critiqued our… · p. 21

Another factor that undermined usefulness and engagement is AI’s insufficient

  • Teachers said AI characters’ formal, everyday-incongruent language undermined authenticity and immersion.They also stressed that appropriate intervention depends on understanding users’ tone and expressing responses suitably in conflict-solving situations.Source: Another factor that undermined… · p. 21
  • A teacher’s example of suggesting a shared PE activity to ease tension illustrated the need for AI to grasp the implications of users’ additional information.The teacher said the AI character did not seem to understand that context.Source: Another factor that undermined… · p. 21

seems he doesn't quite grasp the additional information or implications."

Teachers co-designed an interactive bullying simulation to address the scarcity of role-play programs offering in-situ, personalized feedback. Their feedback highlights opportunities for experiential learning and design priorities, alongside concerns about engagement and authenticity.

seems he doesn't quite grasp the additional information or implications.”

  • Teachers recommended adding gamification to improve engagement, while their feedback also pointed to the need for more interactive, authentic experiences.The system’s lack of direct rewards or punishments was identified as a barrier to long-term use.Source: seems he doesn't quite grasp the… · p. 21
  • Teachers reported that the system offered no direct behavioral rewards or punishments, which they said hindered long-term use, and recommended borrowing ideas from games.Suggestions included rewards tied to increases in joy or trust.Source: seems he doesn't quite grasp the… · pp. 21–22

players could earn something like gems or other rewards.”

  • Teachers saw conversational interaction as encouraging active thinking and participation, while noting animation may better suit younger children who struggle to organize their words.Views differed on which format is more engaging; the authors preferred integrating animation in a future version to improve illustration and cover more age groups.Source: players could earn something like gems… · p. 22

Discussion

  • The study used two design sessions to understand bullying situations in K–12 schools and develop a prototype with conceptualization, experimentation, and analysis parts.The discussion identifies a gap: few programs offer role-play simulation with in-situ, personalized feedback.Source: Discussion · p. 23

5.1 AI as an Efficient Tool to Support Experiential Learning.

  • The authors describe LLM support across experiential learning stages: simulating characters, analyzing students’ prior decisions with summative feedback, and allowing open experimentation with immediate feedback.They state that these processes can help students retain and practice skills.Source: §5.1 · p. 24

5.2 Key design parameters for a bullying simulation system

  • Teachers’ accounts indicated that bullying often takes verbal or social forms, and that realistic agents should reflect bullies’ high social status, academic performance, and motivation to maintain standing.The authors identify these traits and motivations as important to authentic simulation.Source: §5.2 · p. 24

According to the feedback from teachers, victims are often portrayed as "weak"

Teachers’ design feedback highlights how victim and bystander characteristics matter in bullying scenarios. The conclusion describes a co-designed LLM simulation prototype and positive participant feedback.

individuals with lower academic performance, poor social skills, or less favorable physical

Conclusion

  • The study co-designed a prototype contextualizing LLMs for multi-participant bullying scenarios and reflecting Kolb’s experiential learning cycle; participants gave positive feedback and feasible future directions.Source: Conclusion · p. 25
Loading 2609.35776v1…