Bias measurement

Previous research has established an extensive set of metrics and benchmarks for quantifying bias in language models [30]. These methodologies can be broadly categorized into two families. The first approach involves analyzing the models’ weights to their likelihood of generating specific tokens [30]. The second approach treats the models as black boxes, focusing on either global or local properties of the generated text [2]. In this study, we adopt the latter perspective and evaluate the political bias of the model by estimating the stance of arguments it generates.

While many techniques for quantifying bias predominantly rely on targeted questions or text completion tasks, several studies have investigated biases in downstream applications such as classification [31], search [32], item recommendation [33], and task-specific text generation [34,35], often with a particular emphasis on gender biases [36]. However, these studies do not take into account iterated social interactions between agents.

Epistemia in LLMs

Recent literature has formalized the concept of epistemia, referring to the tendency of Large Language Models (LLMs) to mimic knowledge through lexical plausibility rather than through contextual reasoning or normative evaluation when performing judgment-based tasks [37]. Our work is situated within this line of research. Importantly, we emphasize that the notion of “bias” considered here should not be interpreted in its traditional sense, namely as a flaw that can be straightforwardly corrected or mitigated. Instead, it should be understood as a structural and intrinsic characteristic of text-generative systems. This perspective is further supported by recent studies on generative exaggeration [38], which show that, when simulating human behavior, LLMs do not merely reproduce observed patterns but tend to amplify the most salient traits present in their training data. As a result, they introduce systematic distortions that may fail to capture the full variability and complexity of real human behavior. This clarification is essential to properly situate our work within the broader landscape of these systems.

Personas

LLMs can be conditioned on simulated human ‘personas’ characterized by a set of identity or personality traits specified in the model’s prompt [39]. This role-playing exercise is valuable for simulating human interactions that reflect various properties of a target population [19]. Earlier studies have explored personas defined across multiple dimensions, including political orientation [28,40], personality traits [10,41], and sociodemographic characteristics [12].

The effectiveness of personas remains a subject of ongoing debate. On one hand, the use of personas has demonstrated significant potential in accurately replicating the preferences and voting behaviors of entire populations [12]. On the other hand, these personas often carry inherent biases that stem from societal stereotypes associated with them [7]. Furthermore, the influence of personas on certain text-generation tasks appears to be marginal [42]. Similar to our study, previous work showed that when agents are initialized with OCEAN personality traits [43], they generate responses that are more aligned with the assigned personality when queried with one-off prompts than when engaging in a collaborative task with other agents [44]. Our experiments go beyond that and show that agents initialized with a persona may generate outputs typical of an opposite persona when immersed in conversational settings.

Multiagent systems

The rise of LLMs capable of generating human-like dialogue has prompted interest in using generative AI to simulate social systems [27]. Research in this area has focused on multi-agent collaboration [20] and the creation of in-silico societies that mirror real-world dynamics [19,45]. These synthetic populations aim to support social science by emulating diverse human responses [12,46], though concerns remain about their validity and representativeness [47]. In line with these critiques, we argue that LLM biases may distort the realism of simulated debates.

Ashery et. al. [24] study small populations of LLM agents in the theoretical setting of the Naming Game. In each round, two randomly selected agents independently choose a word from a limited inventory; matching choices are rewarded and mismatches penalized. Repeated interactions lead to global consensus via local coordination. Notably, they show that consensus can emerge around words that individual agents are unlikely to select, a phenomenon they call collective bias, where the collective outcome diverges from the individual model bias.

Our experimental design builds on prior work in opinion dynamics among generative agents [48]. Typically, agents communicate over network topologies and update their opinions based on peer input [28]. While prior studies show that agents tend to reject false claims, often due to safety filters [49], we demonstrate that repeated interactions can lead to sudden and complete reversals of an agent’s intended persona.

Experimental setup

We combine two elements to simulate online interactions: agents capable of mimicking human behavior and reasoning, and an environment in which the agents can interact. Next, we detail how we design our social simulation setup and explore how these agents may behave in real online social environments. We depict a sketch of the proposed framework in Fig 1.

Download:

• PNG
  larger image
• TIFF
  original image

Fig 1. Conversational framework.

The social agents take turns generating messages based on the message history. Messages that contain an opinion, according to the opinion presence agent, are passed to the opinion signal agent to estimate the corresponding stance. The messages are real examples from our simulations.

https://doi.org/10.1371/journal.pone.0355458.g001

Chatroom

We simulate a simple chatroom environment defined by: (i) the topic of discussion; (ii) the number of agents ; and (iii) the total number of chatroom messages . In a chatroom, multiple social agents engage in a conversation on a specific topic. Each agent takes turns sending messages while having access to all messages from other participants. This setup effectively represents a fully connected network of agents. Agents are assigned with an initial opinion, and we use the opinion signal agent to evaluate the opinion of an agent each time they generate a new message. For generation, at each time step we randomly select an agent from the set of agents, excluding the one that sent the message at time step . This ensures that no agent generates two consecutive messages. The simulation ends when messages have been generated. Notice that, at the end, the final conversation can still affect the opinion of the agents that did not generate the last message. To track this, we allow every agent except the one that generated the -th message to generate an internal message, i.e., a message that the other agents are not exposed to. We use these internal messages to check if an opinion change occurs after the last message is posted. Thus, for a chatroom with messages and agents, we generate a total of normal messages and internal messages.

Social agents

The agents are initialized with a system prompt (see an example in Box 1) consisting of four elements that closely follow prior work in LLM opinion simulations [48]:

1. An introduction to the conversation setting.
2. The stance of the agent on the conversation topic, spelled as: “You agree with the statement [STATEMENT].”.
3. A longer description of the given opinion.
4. An example statement representing the given opinion.

We consider eight politically polarizing topics sourced from recent polls [50–52]: abortion, climate change, gender identity, gun control, healthcare, immigration, marijuana legalization, and racial attitudes. Statements from these polls, annotated with liberal and conservative agreement rates, guide our prompt design. For consistency, agreement with the original statement reflects a liberal stance, while agreement with its complement reflects a conservative stance. Table 1 provides an example statement for each topic and the actual percentage of Liberal and Conservative citizens agreeing with it.

Download:

• PNG
  larger image
• TIFF
  original image

Table 1. For each topic, an example statement and the actual percentage of Liberals and Conservatives agreeing with it, according to the Gallup Poll and the Pew Research Center.

https://doi.org/10.1371/journal.pone.0355458.t001

The agents take random turns at broadcasting messages to the chatroom and rely on a memory module to keep track of the message history, which is included as context within the prompt. We implement the memory module using LangChain [53], and opt for the list-based memory strategy over summarization-based strategies that are prone to introduce noise. This configuration truncates older text when the context window is exceeded.

Opinion agents

To gauge how the evolution of the agents’ opinion, we must assess the stance expressed by the generated messages over time. In current practice, this is achieved using a stance classifier [48]. We use LLaMa3-70B-Instruct as an opinion signal agent that classifies the stance of each message individually. This is achieved in a few-shot fashion, with five opinion classes and a standard 5-point Likert scale [54].

In our context, we provide the classifier with the statement that represents the topic of the chatroom, paired with one example each. For robustness, we query the opinion signal agent 10 times per message and use the majority label. In case of ties, we repeat the query on the tied options. Since not all messages express an opinion, we first use a similarly configured opinion presence agent to determine whether an opinion is present. This preliminary check prevents the opinion signal agent from being invoked on messages that do not contain any explicit opinion towards the discussed topic, thus avoiding potentially incorrect estimates of unwarranted opinion changes. Finally, the agent’s system prompt is updated with the new opinion, by changing the level of agreement with the topic statement.

To evaluate the accuracy of the opinion signal agent, we manually annotated 1,000 messages with their corresponding stance, following a structured annotation procedure. Topics were distributed among the authors such that each topic was labeled by a single author. We assigned entire topics to individual authors proved to be the most time-efficient approach. This method also helps preserve internal consistency in labeling criteria within each topic. After completing the annotations, we input the messages into the opinion signal agent and compared its predicted labels against our human-provided annotations. We found that the opinion signal agent achieves a macro F1 score of 0.84. This is in line with prior observations on the reliability of LLMs for stance detection [55–58]. In the Appendix, we additionally provide the confusion matrix on the topics of Abortion, Climate Change, Healthcare, and Marijuana Legalization, showing that the model is accurate on both political poles.

To assess the above finding, we further sampled a subset of 200 messages from the 1,000 previously annotated messages and distributed them among all authors for additional manual annotation. The goal was to evaluate the robustness of the macro F1 score in light of potential annotation ambiguity across authors. To this end, we computed Cohen’s Kappa score [59] to estimate inter-annotator agreement. We obtain a score of (substantial agreement) when categorizing labels into “Liberal,” “Conservative,” and “Neutral,” and a further lower score () when distinguishing between Strongly and Slightly Liberal/Conservative. These results further confirm the difficulty of the labeling task and help contextualize the performance achieved by the opinion stance model.

Setup

We empirically evaluate our framework with nine state-of-the-art models as alternatives for the social agents:

• Claude-3.5-Sonnet (https://www.anthropic.com/news/claude-3-5-sonnet): Developed by Anthropic, Claude 3.5 Sonnet is an advanced AI model that excels in reasoning, coding, and safety.
• Gemini-1.5-Pro (https://deepmind.google/technologies/gemini/pro/): A multimodal AI model developed by Google DeepMind, designed to enhance generative AI services across Google’s platforms and for third-party developers.
• Gemma1.1-7B-It (https://huggingface.co/google/gemma-1.1-7b-it): A lightweight decoder-only large language model developed by Google, featuring 7 billion parameters.
• GPT3.5 (https://openai.com/index/gpt-3-5-turbo-fine-tuning-and-api-updates/): An advanced conversational AI model developed by OpenAI, built on the GPT-3.5 architecture.
• GPT-4o (https://openai.com/index/hello-gpt-4o/): Introduced by OpenAI in 2024, ChatGPT-4o is a multimodal LLM capable of real-time conversations, question answering, and text generation.
• LLaMa3.1-70B-Instruct (https://huggingface.co/meta-llama/Llama-3.1-70B-Instruct): A language model developed by Meta, and optimized for multilingual dialogue applications, natural language understanding and generation tasks.
• Nous-Hermes-2-Mixtral-8x7B (https://huggingface.co/NousResearch/Nous-Hermes-2-Mixtral-8x7B-DPO): A flagship language model developed by Nous Research, built upon the Mixtral 8x7B Mixture of Experts (MoE) architecture.
• Qwen2.5-72B-Instruct (https://huggingface.co/Qwen/Qwen2.5-72B-Instruct): A state-of-the-art instruction-tuned language model in the Qwen series, excelling in knowledge, coding, mathematics, and instruction-following capabilities.
• Zephyr-7B (https://huggingface.co/HuggingFaceH4/zephyr-7b-beta): A 7-billion-parameter language model developed by Hugging Face’s H4 team, fine-tuned for tasks such as writing, role-playing, translation, and text summarization.

For each combination of topic, echo chamber type, and language models, we run 50 simulations each consisting of N = 2 agents and M = 20 total messages (10 per agent). Remarkably, we were not able to obtain empirical results with Claude-3.5-Sonnet towards the Liberal pole, since the model safety guards prevented impersonating Conservative personas (the agent keeps generating sentences like “I will not roleplay or express views on controversial political topics”.). The opinion presence and opinion signal models (LLama3-70B) ran on an NVIDIA DGX equipped with 3 V100(32GB) GPUs and 640GB memory. Other models were invoked through the HuggingFace, OpenAI, Anthropic, and Google Gemini APIs. When applicable, it has been used a temperature equal to 0.7. Our code is publicly available (https://github.com/EricaCoppolillo/LLMsConversationalBias).

Ethical considerations

Our analysis leads to important considerations from an ethical perspective. Specifically, the reported findings suggest that the presence of conversational bias may potentially affect the deployment of LLMs in several scenarios, which we address in the following.

1. The safe adoption of LLM agents in social scenarios could be compromised, such as the implementation of bots to be introduced in social platforms (e.g., , Facebook). In particular, the resulting interactions may contribute to amplify societal biases and influence public opinions, thus contributing to the formation and strengthening of echo chambers and polarization or the promotion of harmful and discriminatory content.
2. Even worse, the deployment in education, mental health support or customer service could result in misleading or harmful outcomes. For instance, in the context of mental health support, the agent may exhibit an unwarranted deviation from the original instructions and encourage the user to maintain unhealthy or dangerous habits (e.g., drug addiction, depression).
3. The bias may also affect the studies based on simulated approaches in social contexts (e.g., the evolution of echo chambers, user polarization, opinion drifts). In fact, without proper calibration, the underlying bias may compromise the effectiveness of such studies and consequently their credibility.
4. It is also crucial to recognize that these biases can be deliberately exploited by malicious users. Adversaries can strategically probe an LLM agent within a social context, crafting interactions designed to reveal latent biases that may not be immediately apparent. Once identified, these biases can be manipulated to influence the model’s responses in predictable ways, potentially steering conversations toward misleading, biased, or harmful conclusions. This type of exploitation poses significant risks, particularly in sensitive domains such as politics, social discourse, and automated decision-making, where LLMs may unknowingly reinforce ideological slants or discriminatory patterns.

One-shot bias assessment

We first test the model bias using a state-of-the-art approach based on direct probing [8]. We instantiate a Strongly Liberal and a Strongly Conservative social agent, similar to the chatroom setting, and we query them for their Likert-scale [54] agreement on a statement on a given prompt. Box 1 provides an example, where [EXAMPLE STATEMENT] indicates a representative statement of each political leaning (e.g., “Abortion is a woman’s choice and no one should be able to take it away.” for Strongly Liberal, and “There is no justification for abortion. Any life is important, no matter the circumstance.” referring to Strongly Conservative). We query each agent 10 times and calculate the average Likert deviation . By “Likert deviation”, we refer to the degree to which an LLM response diverges from the expected answer (“Strongly agree”). A deviation of 0 indicates no change, that is, the model consistently answers “Strongly agree” across all 10 trials. Higher values represent greater movement along the Likert scale away from the expected response. The maximum deviation (4) occurs when the model answers “Strongly disagree”, which is four steps away from “Strongly agree”. The + /– signs indicate the direction of the deviation, toward the Liberal or Conservative pole. For instance, if the agent is instantiated as Strongly Liberal but responds in a Strongly Conservative manner (e.g., replying “Strongly disagree” to a prototypical Liberal statement), we assign a score of +4. Conversely, if the agent is instantiated as “Strongly Conservative” and gives “Strongly disagree” to a representative Conservative statement, we assign a score of .

As the left side of Fig 2 shows, this method does not detect strong biases. Most models exhibit responses consistent with the assigned persona, except for Gemma1.1 and GPT3.5 showing a Liberal inclination.

Download:

• PNG
  larger image
• TIFF
  original image

Fig 2. Left: Average Likert Deviation computed via direct probing.

The score represents the average across 10 different trials. Right: computed via our framework. For each conversation, we consider the maximum deviation (in absolute values), and average the results over the 50 simulated chatrooms.

A score of 4 indicates that the LLM instantiated as Liberal responded in a strongly Conservative manner, while indicates the opposite. Empty cells indicate 0-values, while cells with the “-” symbol indicate unavailable results. The darker the color of the cell, the more prominent the detected shift.

https://doi.org/10.1371/journal.pone.0355458.g002

Conversational bias assessment

To assess the validity of our framework, we recompute the average Likert Deviation by applying our multi-agent conversational setting (Fig 2, right). Specifically, we compute for each conversation the maximum deviation detected (in absolute value) and average the results on the 50 simulated conversations. Our methodology reveals much more frequents unwarranted drifts than direct probing, especially from the Conservative toward the Liberal pole. In the Appendix, we further provide the percentage of dyadic chatroom simulations where at least one unwarranted opinion change occurs.

Crucially, if we compare left and right side of Fig 2, we see that, contrary to the one-shot bias assessment, all models exhibit a strong and systematic conversational bias. Notably, this bias is predominantly towards Liberal stances, with occasional shifts from Liberal towards Conservative. For instance, while Llama3.1 and Qwen2.5 exhibit the greatest shift on “Marijuana legalization”, ChatGPT-4o shows the most prominent bias towards “Healthcare”. However, also other topics trigger significant bias, such as “Abortion” (Qwen2.5), “Racial Attitude” (LLaMa3.1), and “Climate change” (ChatGPT-4o). Further results showing the agents drift also towards neutral stance are provided in the Appendix. As a side not, we further try different visualizations by modifying the model ordering, to assess potential relationship between the model structural characteristics and the observed phenomenon. Specifically, we sort them according to the parameter count and the context-window side. However, such analysis did not reveal any significant patterns.

Further, we investigate the number of agents changing opinion during the conversation. Specifically, left side of Fig 3 reports the average number of agents displaying an unwarranted opinion. We see that, independently of the topic, at least one Conservative agent always changed opinion towards the Liberal pole. Despite the opposite being rare, in all the configurations also one Liberal agent exhibits a drift over at least one topic. However, the most significant drift remains left-oriented. The percentage of chatroom where one or both agents exhibit an opinion change is depicted in the Appendix. A complementary perspective consists in disentangling the intrinsic bias of the model from potential sycophancy effects. To this end, we estimate the conditional probability that the second agent follows once the first has already drifted. We report the results in the right side of Fig 3.

Download:

• PNG
  larger image
• TIFF
  original image

Fig 3. Left: Average number of agents changing opinion during the conversation.

Right: Conditional probability that the second agent follows once the first has already drifted.

Empty cells indicate configurations where no agent displayed a drift, while cells with the “-” symbol indicate unavailable results. The darker the color, the higher the reported value.

https://doi.org/10.1371/journal.pone.0355458.g003

Sensitivity analysis

To test the robustness of our findings, we perform different variations of the experimental setup by modifying the length of the conversations, ablating the prompt, and increasing the number of conversational agents.

Conversation length. First, we investigate how the conversation length affects the emergence of the bias. Specifically, we aggregate the results in terms of unwarranted opinion change across all topics for a given LLM, and we analyze how varying the value affects the corresponding percentage. We depict the results in Fig 4, where the left (resp. right) subplot shows the percentage of Liberal (resp. Conservative) conversations having an unwarranted opinion change. Error bars indicate confidence intervals. Two important considerations emerge: (i) for most models, the drift occurs even after a few messages between the agents; and (ii) this pattern is far more pronounced when the agents are instantiated with conservative personas. In the Appendix, we further report the same percentages computed at the topic-level.

Download:

• PNG
  larger image
• TIFF
  original image

Fig 4. Percentage of Liberal (left) and Conservative (right) conversations exhibiting an unwarranted opinion change (Y-axis) by varying the conversation length M (X-axis).

Error bars indicate confidence intervals.

https://doi.org/10.1371/journal.pone.0355458.g004

Prompt ablation. To rule out the possibility that observed biases were artifacts of prompt design, we conducted a series of robustness checks using multiple prompt variants. These included removing contextual cues such as “you are participating in a chat” and introducing stronger identity reinforcement through statements like “you are really protective of your view about [TOPIC]”. Across all variations, the results remained consistent, indicating that the bias is not sensitive to minor prompt modifications. Additionally, while agent names were initially assigned following conventions from prior work [44], we performed an ablation study removing agent names altogether. This had no measurable impact on the outcomes, further confirming that the observed bias is not driven by superficial prompt-level cues.

Number of agents. Finally, to assess the scalability of our findings, we extended the experimental setup to conversations involving larger groups of agents. As representative candidates, we tested Mixtral on the “Climate Change” topic with [2,5,10]. We find no bias towards the Conservative pole, while devising drifts toward the Liberal pole in 92%, 96%, and 86% of the conversations, respectively. All the observed results were statistically significant. This indicates that increasing the number of agents does not significantly affect the emergence of bias. This follows naturally from the design of our framework, in which each agent is initialized with a stance via a system prompt and generates responses based on the evolving conversation history. While increasing the number of agents lengthens this history, the bias phenomenon remains observable even in short interactions, as shown in Fig 4. This supports the generalizability of the findings from dyadic to multi-agent settings.

Memory loss. Next, we investigate whether the observed bias can be due to memory loss. Specifically, we verify the saturation level of the context window of the given model L, which we denote as . Such information is publicly available for the tested models, and is reported in Table 2. To assess if saturation occurred, we consider all the 50 generated conversations for each combination of agent and topic. We further tokenize each conversation to assess the longest produced by a given model across all the topics. For each model L, we denote the number of tokens of such longest conversation as . To ensure robustness, we perform the tokenization task by using two different tools: a BERT-based model tokenizer from HuggingFace (https://huggingface.co/docs/transformers/main_classes/tokenizer) and tiktoken, a crafted GPT-tokenizer released by OpenAI (https://pypi.org/project/tiktoken/). If the amount of estimated tokens differs, we consider the highest. To assess if a memory loss occurred, we therefore compute , representing the portion of memory actually occupied by the longest conversation ever produced. The results are provided in Fig 5. As shown, the highest percentage of memory ever occupied is about 64%, when Zephyr is instantiated as agent, followed by Mixtral (30%). We hypothesized this is due to the relatively small window context capacity of these models (see Table 2) combined with their average amount of tokens produced, displayed in Fig 6. Notably, this analysis shows that no agents display memory loss during the simulated conversations, thus excluding this factor as cause of the observed bias.

Download:

• PNG
  larger image
• TIFF
  original image

Table 2. Context window size for each tested model, expressed in number of tokens.

https://doi.org/10.1371/journal.pone.0355458.t002

Download:

• PNG
  larger image
• TIFF
  original image

Fig 5. Portion of context memory occupied by the longest conversation ever produced by a given model across all topics.

https://doi.org/10.1371/journal.pone.0355458.g005

Download:

• PNG
  larger image
• TIFF
  original image

Fig 6. Average number of tokens, expressed in thousands (K), generated by each model across all topics.

Error bars represent standard deviations.

https://doi.org/10.1371/journal.pone.0355458.g006

Opinion signal agent. Finally, we assess the robustness of our results by varying the underlying opinion signal agent. Specifically, we adopted Gemini-2.0-flash (https://blog.google/technology/google-deepmind/google-gemini-ai-update-december-2024/), released by Google in 2024 with improvements in multimodal reasoning and long-context understanding. We use the conversation generated by Mixtral on “Climate Change”, both prompted as Liberal and Conservative, as a representative sample. For a fair comparison, we prompted the opinion signal agent with the same protocol reported in the paragraph “Opinion Agents”.

The results are reported in Table 3. The first (respectively, second) row shows the percentage of conversations with conservative- (respectively, liberal-) prompted agents in which a drift toward the Liberal (respectively, Conservative) pole was detected. The last row presents the aggregated results across all conversations. The column “Agreement” reports the percentage of conversations where the opinion signal agents agree in terms of drift (detected/non-detected).

Download:

• PNG
  larger image
• TIFF
  original image

Table 3. Percentage of conversations where an opinion drift was detected by any of the two opinion signal agents. The first (resp. second) row refers to conversations where conservative- (resp. liberal-) agents interacted. The last row depicts the aggregated results across all conversations. The column “Agreement” reports the percentage of conversations where both the opinion signal agents agree in terms of detected drift.

https://doi.org/10.1371/journal.pone.0355458.t003

Two key observations emerge. First, the two opinion-signal agents demonstrate an 84% agreement across the 100 selected conversations (50 per political leaning). Second, Gemini-2.0-Flash is more likely than LLaMa-3-70B to flag conversational drifts, labeling a larger portion of content as shifting toward the opposite pole. These patterns imply that the results reported in the previous sections remain valid even under a majority-voting scheme, since the conversations where both agents agree coincide with those flagged by LLaMa-3-70B. In other words, LLaMa-3-70B effectively provides a lower-bound estimate of the proportion of simulations exhibiting an unwarranted opinion drift.

Limitations

This work presents several limitations that open avenues for future research and refinement.

1. The study centers primarily on polarizing U.S. political topics, such as climate change and abortion [63,64]. While these are salient and well-documented issues, this focus may limit the generalizability of findings to less contentious or non-U.S. contexts, where the structure of opinion formation and political engagement may differ significantly.
2. The reliance on an automated stance detection system introduces an additional layer of abstraction and complexity. Although the used model proves to accurately perform the task, it may introduce potential inaccuracies and biases inherent to the stance classifier itself. An alternative approach could involve endowing conversational agents with the ability to explicitly signal their positions, increasing interpretability and reducing dependence on post-hoc classification tools.
3. The scope of the work is currently limited to simulations of opinion dynamics within multiagent interactions. While valuable, this narrow task focus may not capture the full spectrum of challenges in human-AI communication. Future work should explore how similar dynamics unfold in other collaborative or adversarial multiagent scenarios, such as joint decision-making, negotiation, or misinformation detection.
4. The range of agent roles modeled in the simulations is relatively constrained. Agents are treated largely as homogeneous participants, lacking more nuanced behavioral archetypes (e.g., skeptics, persuaders, or contrarians). Incorporating diverse agent personas could lead to richer dynamics and a better approximation of real-world conversational ecosystems.
5. The framework has been tested on a comprehensive but limited set of underlying language models. Expanding the evaluation to include a broader range of models, including those with varying capacities, training data, and architectural inductive biases, could provide deeper insights into the generalizability and robustness of the observed phenomena.