Generative Agents for Multimodal Controversy Detection

1 Introduction

Social video platforms serve as important channels for the dissemination of contemporary information, hosting a large amount of video content that spreads worldwide at unprecedented speed [Newman et al., 2023; Gan et al., 2023b]. This content profoundly influences public cognitive perceptions, emotional tendencies, and societal behavior patterns [Wang et al., 2022; Gan et al., 2023a]. The rapid spread and significant impact of large amounts of videos highlight the urgent need for early detection and effective management of dissemination risks [Hessel and Lee, 2019; Wang et al., 2023c]. Particularly concerning are videos that can provoke widespread social controversies [Conover et al., 2011], evoke negative emotions [Wang et al., 2023a], or exacerbate social divisions [Tarrow, 2008]. The identification and management of controversial content, referred to as controversy detection [Martin, 2014], requires increased attention and robust methodologies.

∗Corresponding author.

Multimodal controversy detection is an emerging field that aims to assess whether a given video and its associated content are controversial [Xu et al., 2024]. If not properly managed, controversial videos can spread widely, potentially damaging individual reputations, disrupting group dynamics, and threatening societal stability [Mejova et al., 2014]. Effective controversy detection facilitates timely identification of risks, supports the implementation of appropriate intervention measures, and helps maintain social media platforms as constructive environments for information exchange [Garimella et al., 2018]. Controversial videos on social media can be defined from three perspectives. First, it considers whether the video itself is controversial, such as through sensationalism or violence. Second, it examines conflicts between the video content and user comments, including opposing viewpoints, personal attacks, and criticisms. Third, it focuses on the controversy within the comments, looking for clear support or opposition.

A significant limitation of existing methods is that they only provide a binary output indicating whether a video and its associated content are controversial, without offering detailed explanations. Emphasizing the ethical implications and explainability of model predictions is essential, particularly in sensitive contexts such as controversy. As the adage goes,

"There are a thousand Hamlets in a thousand people's eyes", reflecting the notion that each individual has their interpretation or perspective, which is particularly relevant in understanding controversy. To address this gap, we strive to explore more explainable approaches for this task.

Recent advancements in Large Language Models (LLMs) have significantly enhanced capabilities in natural language understanding and generation [Bang et al., 2023; Jiao et al., 2023]. Various training paradigms have also emerged, enabling LLMs to perform tasks in a zero-shot manner and adhere more closely to human-provided instructions [Sanh et al., 2022]. Although a single powerful LLM is already capable of addressing a wide range of tasks, recent studies suggest that multiple LLMs can further enhance each other's performance through debate and collaboration [Du et al., 2024; Liang et al., 2024]. By integrating multiple LLMs into a cohesive group and designing specific interaction mechanisms, these models can propose and deliberate on unique responses and thought processes across several rounds. Additionally, these agents can simulate human behavior based on userspecified descriptions and profiles [Shanahan et al., 2023]. Such advancements enable LLM-based agents to effectively handle complex scenarios on social media, simulate controversy formation, and improve the explainability of controversy detection. As illustrated in Figure 1, traditional Multimodal Controversy Detection (MCD) models can only output a binary indicator of controversy based on the input video and text. In contrast, incorporating LLM-based agents enhances the models' ability to address more explanatory questions, such as "What is controversial?", "Why is it controversial?", and "How controversial is it?".

In this light, we propose a framework called Agent-based Multimodal Controversy Detection (AgentMCD), which utilizes LLM-based agents to generate explainable text and simulate the formation of controversy. Our approach considers two scenarios: one with abundant comments and another with limited comments. In scenarios with abundant comments, AgentMCD primarily uses LLM-based agents as judges to conduct multi-aspect reasoning, delivering a comprehensive evaluation of video controversies. This includes assessing: (1) the intrinsic controversy of the video content, (2) the controversy between the video content and user comments, and (3) the controversy within the comments themselves. To emphasize the importance of early detection, we also investigate the user engagement simulation in the initial stages of video release, marked by a limited number of comments, where generative agents simulate audience comments and discussions. We evaluate our method using the MMCD dataset [Xu et al., 2024] and compare it with various publicly available LLM-based methods. The experiments demonstrate that AgentMCD achieves notable reasoning performance and effectively simulates controversy with multi-agent mechanism. Ablation studies validate the effectiveness of the proposed multi-aspect reasoning and multi-agent simulation. Finally, we conduct comprehensive case studies and in-depth analyses to assess the effectiveness of the proposed method. This work tackles the critical societal challenge of risk control in short videos, aiming to provide early warnings and directional guidance for risks encountered during the dissemination of short videos.

Our main contributions are summarized as follows:

• We are the first to explore incorporating LLM-based agents into the task of multimodal controversy detection, thereby enhancing the explainability of the process.
• We introduce a novel framework called Agent-based Multimodal Controversy Detection (AgentMCD), which utilizes a multi-aspect reasoning process to systematically evaluate controversy. Furthermore, it incorporates a multi-agent simulation mechanism to model the formation of controversy in the early stages of video dissemination.
• Extensive experiments highlight the superiority of our approach compared to other LLM-based methods, while maintaining a good explainability. Our work is publicly available 1 .

2 RelatedWork

2.1 Controversy Detection

Controversy detection has garnered significant attention in recent research due to its broad implications across various domains such as risk management, content moderation, and sentiment analysis [Linmans et al., 2018; Zhong et al., 2020; Hessel and Lee, 2019; Xu et al., 2024]. The early studies in this area typically assumed that topics inherently possessed controversy and focused on identifying such controversial topics [Popescu and Pennacchiotti, 2010; Garimella et al., 2018]. However, controversial topics cover a wide range of evolving subjects and they do not directly reveal the inherent properties of controversy. Recent studies have further deepened the understanding of controversy identification and analysis, considering various factors such as semantic [Linmans et al., 2018], viewpoint consistency [Hessel and Lee, 2019; Zhong et al., 2020], contextual content [Beelen et al., 2017], and the influence of cultural and social backgrounds [DoriHacohen and Allan, 2015]. The application of controversy detection has expanded to a variety of platforms, from traditional web pages [Dori-Hacohen et al., 2016; Linmans et al., 2018] to more dynamic and socially interactive environments like social media [Garimella et al., 2018; Koncar et al., 2021]. Notably, there has been a growing interest in detecting controversy in multimodal contexts, particularly on social media platforms [Xu et al., 2024], underscoring the importance of addressing the complexities of multimodal content.

To effectively address evolving applications, various methods for controversy detection have been developed. Traditional approaches have primarily relied on statistical techniques [Popescu and Pennacchiotti, 2010; Hamad et al., 2018], which identify controversial content based on specific terms, phrases, or predefined rules. More recent research has shifted towards graph-based methods [Mendoza et al., 2020; Benslimane et al., 2023], which effectively capture structural relationships among users, topics, and comments [Benslimane et al., 2023; Zhong et al., 2020; Li et al., 2023]. Early prediction of controversy has also gained increasing attention [Hessel and Lee, 2019]. For multimodal controversy detection, methods integrating neural network modules have been developed to identify controversial videos based on specific definitions of controversy [Xu et al., 2024]. The advancement of pretrained language models offers new tools and methodologies for controversy detection [Calvo Figueras et al., 2023; Canute et al., 2023]. However, these methods typically provide only binary outputs and lack explanatory details. To address this limitation, our approach incorporates generative agents to deliver more explainable results.

1 Codebase is available at https://github.com/skylie-xtj/ AgentMCD Release.

2.2 Generative Agents

Recent advancements in Large Language Models (LLMs) have showcased their exceptional capabilities in understanding and reasoning in cross-disciplinary research [Bang et al., 2023; Jiao et al., 2023]. A notable development in this field is the emergence of generative agents, which are designed to role-play specific characters or perform targeted tasks [Park et al., 2023; Shanahan et al., 2023]. Researchers have invested substantial efforts in exploring how to utilize these models more effectively to solve various complex problems, thereby advancing the exploration of the application boundaries and potential of LLMs [Kojima et al., 2022; Wang et al., 2023b; Sun et al., 2023; Shinn et al., 2023]. In the field of multiagent systems, current research primarily focuses on simulating human behavior and tackling intricate tasks through collaborative approaches [Park et al., 2023; Zhu et al., 2023; Wu et al., 2023; Hong et al., 2024]. For instance, generative agents [Park et al., 2023] create a sandbox environment that facilitates reliable human behavior simulation, allowing intelligent agents to interact with one another. Ghost in the Minecraft [Zhu et al., 2023] integrates LLMs with text-based knowledge and memory to develop generally capable agents within the Minecraft environment. AutoAgent [Chen et al., 2024] serves as a cooperative agent framework that enables role-playing, collaboration, and the resolution of complex tasks. These methodologies significantly enhance decisionmaking efficiency and accuracy [Zhang et al., 2024].

Significant progress has also been made in simulating social dynamics using LLMs [Cheng et al., 2023; Liu et al., 2023; Liu et al., 2024]. CoMPosTCA [Cheng et al., 2023] highlights LLMs' unique advantages in portraying human traits and managing comic material with personalization and exaggeration. The stable alignment method [Liu et al., 2023] allows LLMs to learn from simulated social interactions, while [Liu et al., 2024] used LLMs to simulate the dynamics of fake news propagation in social media environments and explore its impact on public attitude changes. These studies demonstrate the potential of using LLM agents to emulate human social behaviors, offering new insights into understanding and predicting human actions.

To the best of our knowledge, our work is the first to propose the use of generative agents specifically for multimodal controversy detection.

3 Methodology

In this section, we introduce AgentMCD, a hybrid framework that integrates a controversy simulation mechanism with an explainable reasoning process. This framework is designed to model audience behavior on social video platforms effectively and to predict the long-term engagement impact of videos reliably, even in early-stage scenarios with limited user comments. This is achieved by leveraging the humanlike capabilities of LLM-powered generative agents, providing interpretable outputs for multimodal controversy detection. To realize these objectives, two core components are emphasized: (1) User Engagement Simulation: This component faithfully replicates users' personalized preferences and simulates the process of browsing social videos and commenting. (2) Multi-Aspect Reasoning: This component evaluates controversy scores through a comprehensive process while ensuring good explainability. An overview of the AgentMCD framework is presented in Figure 2, and the complete algorithmic process is described in Algorithm 1.

3.1 Task Formulation

Multimodal controversy detection task aims to detect whether a given video and its associated content are controversial. For an input sample that includes video V and additional context information C, let ˆ y ∈ {0, 1} indicate the predicted result of whether the input is controversial. Our objective is to design an explainable multi-modal controversy detection framework F:

where O denotes the explainable output report, including controversy scores and judgment bases.

3.2 Initialization

The Initialization module outlines the formation of the video profile and user profile within AgentMCD, both of which are fundamental to the subsequent analysis.

Video Profile

In the subsequent module, the video profile-especially the generated video description-plays a pivotal role. Each input dataset includes raw video, title, keywords, metadata about the publishers, comments, and recognized Automatic Speech Recognition (ASR) content. These components, along with the video description generated by the LLM, collectively form the video profile. In general, video descriptions are generated by encoding the input video using a video model aligned with a language model, and subsequently deriving outputs from the language model. Given that the video content is in Chinese, we utilize mPLUG-Video [Xu et al., 2023], fine-tuned on a 10-million Chinese large-scale videotext dataset, to generate descriptions from raw video.

However, this approach occasionally produces irrelevant descriptions. For instance, when publishers express personal opinions on certain topics while appearing on camera, the generated description might merely state, "a person speaking to the camera, occasionally covering their nose or pacing back and forth." Such descriptions are insufficient for our purposes, as we require more detailed and contextually relevant content. Specifically, we are interested in understanding what the person is discussing, the opinions they express, and the conclusions they draw. To address this issue, we further refine the video description through prompt engineering, incorporating additional information from the video profile, including title, keywords, publisher information, and ASR text, to enrich the level of detail in these descriptions.

User Profile

In the subsequent simulation process, the user profile module plays a pivotal role in aligning agent behavior with genuine human responses. Traditionally, agent descriptions are either predefined or randomly selected from an established dataset [Zhang et al., 2024; Park et al., 2023; Liu et al., 2024]. However, we observed during experiments that such initialization often led to repetitive responses and a lack of diverse perspectives.

To overcome this limitation, we explored alternative initialization strategies. Drawing inspiration from the personalized recommendation mechanisms commonly used on video platforms, we began generating audience profiles based on video descriptions. We also refined the prompting process to encourage a broader range of opinions.

Each agent's profile consists of two key components: social role and core perspective. The social role includes characteristics such as profession or personal interests-for example, a student preparing for a teaching certification. The core perspective captures the agent's main concerns or viewpoints; for instance, a media commentator may approach a case from a news angle, focusing on its societal impact and media coverage. This design significantly improved the diversity of generated opinions, better capturing the distinct social dynamics associated with different videos.

3.3 Simulation

The Simulation module employs a multi-agent mechanism to replicate the user engagement process, enabling the observation of controversy emergence in complex social video platform scenarios and offering valuable insights into the dynamics of controversy. This module is particularly crucial in contexts where commentary is sparse.

Generative agents in AgentMCD, built on an LLM-based foundational architecture, are enhanced with two specialized modules designed specifically for social video platform scenarios: the profile module and the action module. To emulate personalized and authentic human behavior, each agent incorporates a user profile module that combines social roles and core perspectives, effectively representing the types of users likely to view the video and the aspects they might focus on. Additionally, inspired by the processes humans undergo while interacting with social video platforms, the agents are equipped with action modules that enable them to comment and express emotions in a coherent and contextually appropriate manner. They comment after viewing videos, observe other audiences' reactions and comment, simulating the reactions of viewers encountering the video on a social platform. During these processes, agents' responses may be positive or negative, aiming to replicate the natural flow of interaction between videos and their audiences. This approach effectively simulates the potential conflicts and controversies that may arise from these interactions.

3.4 Reasoning

The Reasoning module serves as the core component of our architecture, where we implement a multi-aspect evaluation Algorithm 1 AgentMCD: Agent-based Multimodal Controversy Detection

• 1: Input: Video title T, commentsC, publisher information P, ASR text R, threshold t
• 2: Output: Predicted controversy label ˆ y, output report O
• 3: Generate detailed video descriptions D based on T, P, R
• 4: if C is None then
• 5: Simulation Process:
• 6: Generate adaptive agents' descriptions A based on D
• 7: for each agent description a in A do
• 8: Initialize Agent i based on a
• 9: Generate comments on D and add to C
• 10: Generate more comments based on C and add to C
• 11: end for
• 12: end if
• 13: Multi-aspect Reasoning Process:
• 14: Aspect 1: Evaluate controversy of video content based on D
• 15: Aspect 2: Evaluate controversy between video D and comments C
• 16: Aspect 3: Evaluate controversy within comments C
• 17: Judge the final controversy score s based on the evaluations of the above three aspects
• 18: Final controversy score s and judgement basis b are added into O
• 19: ˆ y ← {0, if s <t, 1, otherwise.
• 20: Return: Predicted controversy label ˆ y, output report O

process to comprehensively assess controversy from three dimensions: (1) controversy inherent in the video itself, such as the presence of sensationalism, violent content, or other risky elements; (2) controversy between the video and its associated comments, where videos are considered controversial if comments reflect opposing viewpoints, include personal attacks on the video creator, critique the phenomena depicted, or question the individuals or objects featured; and (3) controversy among comments, characterized by disagreements, such as debates or clear expressions of opposing viewpoints (e.g., support versus opposition).

In scenarios with limited comments, we incorporate generated comments from the Simulation module to supplement the evaluation. The final prediction is derived by synthesizing the assessments across these three dimensions. Each aspect of evaluation includes two key components: (1) controversy score, ranging from 0 to 9 (with 0 indicating no controversy and 9 representing high controversy), and (2) judgement basis, which explains the assigned score by identifying the specific sources and reasons for the controversy.

4 Experiments

In this section, we present experiments to compare our proposed models with various baseline models. Specifically, we aim to address the following evaluation questions:

EQ1: Is our proposed AgentMCD framework more effective than other LLM-based methods in scenarios?

• EQ2: Does our proposed framework effectively utilize multi-aspect reasoning, and does multi-agent demonstrate effectiveness?
• EQ3: How does the quality of comments generated by the simulation module compare to real comments?
• EQ4: How does the multi-aspect reasoning module enhance the interpretability of multimodal controversy detection tasks?

4.1 Dataset

We evaluate the MMCD dataset [Xu et al., 2024], a largescale Multimodal Controversial Dataset consisting of over 10,000 Chinese videos, accompanied by extensive social context information. The MMCD was sourced from Douyin 2 , a popular Chinese social video platform with a substantial user base comprising millions of active participants. For our experiments with LLM-based methods, we utilize only the validation and test data, excluding the training data. Specifically, the validation set contains 1,130 samples, while the test set includes 1,132 samples.

4.2 Implementation Details

Our proposed framework is implemented using a Python script, utilizing GLM4-9B [GLM, 2024] as the backbone LLM with a temperature setting of 0 to ensure reproducibility. GLM4-9B, from the latest GLM-4 series by Zhipu AI, supports long-text reasoning up to 128K tokens. All experiments are conducted in one-shot or few-shot settings without additional training or fine-tuning of the language models. Each method first determines the optimal threshold based on the validation set, which is then used to evaluate the test set results.

4.3 Baselines

To validate the effectiveness of our methods, we implemented several representative LLM-based methods on the MMCD dataset. The following elaborates on them:

Standard Prompting employs the GLM4-9B [GLM, 2024] backbone model with a standard zero-shot prompt. Zero-shot Chain of Thought (CoT) [Kojima et al., 2022] uses "Let's think step by step" as a prompt for LLMs. Plan-and-Solve (PS) [Wang et al., 2023b] involves devising and executing a task plan in two steps. Self-Consistency [Wang et al., 2023d] samples multiple LLM responses and uses majority voting to determine the final answer. Self-Reflect [Shinn et al., 2023] involves LLMs refining their outputs until satisfactory. SelfRefine [Madaan et al., 2023] improves initial LLM outputs through iterative feedback. Tree of Thoughts (ToT) [Yao et al., 2024] enables decision-making by evaluating reasoning paths and adjusting actions. Cumulative Reasoning (CR) [Zhang et al., 2023] uses iterative, cumulative reasoning to solve problems. RECITE [Sun et al., 2023] retrieves relevant passages from LLM memory before producing answers.

We also incorporate several multi-agent methods: AutoAgents [Chen et al., 2024] is a framework for building and coordinating AI agents. Multi-Agent (Debate) [Du et al., 2024]

2 https://www.douyin.com/

Table 2: Experimental results of the ablation study, including F1score, recall, precision, and accuracy.

involves multiple LLM instances proposing, debating, and refining answers. MAD [Liang et al., 2024] uses a debate process with multiple agents and a judge to reach a final solution.

4.4 Main Results (EQ1)

Table 1 presents the quantitative evaluation, where AgentMCD consistently outperforms other LLM-based approaches in both abundant and limited comment scenarios, demonstrating its superior effectiveness and simultaneously achieving explainability.

In scenarios with abundant comments, LLM-based methods such as Zero-shot CoT [Kojima et al., 2022], Plan-andSolve [Wang et al., 2023b], Self-Reflect [Shinn et al., 2023], and Tree of Thoughts [Yao et al., 2024] exceed the performance of the Standard Prompting approach. However, some previous multi-agent methods, such as AutoAgents [Chen et al., 2024], Multi-Agent (Debate) [Du et al., 2024], and MAD

[Liang et al., 2024], have not performed well. These methods rely on cooperative mechanisms to achieve objectives or debate mechanisms to refute opposing arguments and reach a correct answer. In contrast, AgentMCD introduces an innovative framework incorporating a multi-agent simulation mechanism that specifically models user engagement within the video propagation process and simulates the formation of controversies. In scenarios with limited comments, not only do multi-agent methods but also certain LLM-based methods that perform well in abundant environments fail to enhance the performance of Standard Prompting. Thus, AgentMCD proves to be particularly effective in such environments by simulating real-world controversy formation.

4.5 Ablation Studies (EQ2)

We conducted a series of ablation experiments to evaluate the significance of various components in the proposed AgentMCD framework, with results presented in Table ??. Specifically, we examined the contribution of each aspect within the reasoning module by independently evaluating controversy within video content (Aspect 1), controversy between video content and comments (Aspect 2), and controversy among comments (Aspect 3). The results demonstrate that each aspect plays a vital role in the overall performance of AgentMCD, with Aspect 3 consistently achieving the best performance across all scenarios. Furthermore, the ablation study revealed that AgentMCD with a single agent underperforms compared to its multi-agent counterpart, underscoring the effectiveness of the multi-agent mechanism.

4.6 Simulation Evaluation (EQ3)

To evaluate the quality of generated comments in the simulation module, we employed an additional LLM (the Qianwen series) alongside human evaluators to score both real and generated comments across multiple dimensions. The results are presented in Table ??. The evaluation criteria include: (1) Relevance: examining topic alignment; (2) Consistency: evaluating sentence coherence; (3) Logicality: measuring the logical flow; (4) Sentiment Polarity: determining emotional Aspect 1: Controversy of video

Controversy Score: 5

Judgment basis: The video content involves a labor controversy, which may spark public discussions on the protection of labor rights and the rationality of company policies. At the same time, the event depicted in the video may lead to consumer dissatisfaction with the supermarket's management practices and raise concerns about employee rights. Additionally, the high amount of the fine in the video could be seen as an infringement on personal rights.

Aspect 2: Controversy between video and comments

Controversy score: 7

Judgment basis: There is widespread skepticism and criticism of the supermarket's regulations in the comments, with sympathy expressed for Ms. Zhang's situation and dissatisfaction with the supermarket's punitive measures. At the same time, some commenters offered differing views on the employee's behavior and the company's policies. Strong emotional expressions, such as anger, disappointment, and condemnation, as well as opinions from legal experts regarding the incident, all contribute to the controversy surrounding the video content.

Aspect 3: Controversy among comments

Controversy Score: 8

Judgment basis:

1. There is a clear opposition of opinions in the comment s . Some suppor t the supermarket's regulations, arguing that employees should follow the rules, while others side with the employees, claiming the supermarket's punishment is too harsh.
2. Emotional expressions appear in the comments, such as "What kind of r idiculous rule is this?", "This is outrageous!", and "This company really looks down on people," showing the commenters' highly emotional reactions to the incident.
3. Some commenters raised legal points, a rgu i ng t ha t t he s upe rma r ke t ' s punishment might violate labor laws, indicating a discussion of legal issues in the comments.
4. Some commenters called for media attention to the incident and urged industry regulations, suggesting a deeper concern with the underlying issues behind the event.

Overall evaluation

Controversy Score: 8

Judgment basis:

1. The video content directly involves labor rights and company regulations, topics that are inherently sensitive and prone to provoke public attention and discussion.
2. The high fine amount in the video is directly related to individual rights, which may lead to public questioning of the fairness of the punishment.
3. There is a clear opposition of opinions in the comments, including differing views on supermarket regulations, employee behavior, and company policies, which intensifies the controversy.
4. Emotional expressions and strong emotional reactions, such as anger, disappointment, and condemnation, indicate that the controversy has touched a sensitive nerve in the public.
5. The involvement of lega l exper t s and discussions of legal issues add professionalism and depth to the controversy.
6. Some commenters have called for media attention and industry regulations, showing that the controversy has sparked broader attention and reflection.

Table 3: Comparison between real and generated comments across various aspects, including relevance, consistency, logicality, sentiment polarity, and diversity, presented through both LLM-based evaluations and human assessments.

intensity; and (5) Diversity: assessing the richness of perspectives. Results suggest that generated comments generally outperform real ones, especially in logicality. Case analysis revealed that real comments often scored lower due to noise, topic deviation, brevity, incomplete arguments, insufficient analysis, and a lack of causal reasoning. Conversely, while generated comments excelled in all dimensions, they lacked the authenticity found in the imperfections of real comments.

4.7 ACase Study (EQ4)

Figure 3 presents a case study of a controversial video, detailing the evaluation results in the reasoning module. For the controversy of video aspect, the content potentially involves a labor controversy. In the controversy between video and comments aspect, it highlights critical comments questioning the supermarket's regulations or expressing sympathy towards customers, accompanied by strong emotions. Regarding controversy among comments, it detects heated exchanges with clear support and opposition. A comprehensive summary of all three aspects is also provided.

The comprehensive output report enhances the interpretability of multimodal controversy detection tasks and addresses the three key questions posed in Figure 1: (1)"What is controversial?": The evaluation reveals that the controversy scores for Aspect 2 and Aspect 3 are 7 and 8, respectively, indicating that the controversy arises from the interaction between video content and comments, as well as among the comments themselves. (2)"Why is it controversial?": This is explained through the judgment basis provided in the evaluation results. (3)"How controversial is it?": The degree of controversy is quantified by the controversy scores in the evaluation results.

5 Conclusion

We propose AgentMCD, a comprehensive framework that leverages a multi-agent mechanism to simulate user engagement and the formation of controversies, along with a multiaspect reasoning process to assess and interpret these controversies. This study explores the application of LLMs to understand the emergence of controversies during video dissemination, enhancing the explainability of traditional multimodal controversy detection tasks. Empirical evaluations reveal that AgentMCD significantly outperforms existing LLM-based approaches. Moreover, the approach of roleplaying to enhance interpretability, as discussed in this paper, can be applied to other domains, such as fake news detection, as well as scenarios involving collaboration and competition.

Ethical Statement

As discussed, LLMs may occasionally produce irrelevant or harmful outputs, necessitating caution when interpreting their results. In our approach, LLM-based multi-agent systems are employed solely to enhance the simulation of controversy formation. However, additional research is required for language models intended for practical applications to refine prediction accuracy and bolster the model's authenticity and safety, thereby mitigating potential user risks.

Acknowledgments

This work is supported by the National Natural Science Foundation of China, No. 62176137; SMP-IDATA Open Youth Fund, SMP2023-iData-007; Fundamental Research Funds for the Central Universities, NO. NJ2024029.