Paper deep dive

Agent Sociodrama: Social-Dynamics-Guided Persona Recasting for LLM Agent Teams

Yirui Qi, Xiaoming Zhang, Jaehong Yoon

Year: 2026Venue: arXiv preprintArea: cs.AIType: PreprintEmbeddings: 78

Intelligence

Status: not_run | Model: - | Prompt: - | Confidence: 0%

Entities (0)

No extracted entities yet.

Relation Signals (0)

No relation signals yet.

Cypher Suggestions (0)

No Cypher suggestions yet.

Abstract

Abstract:Large language model (LLM) agents are increasingly deployed as teams, yet they fail in characteristically social ways: information holders stay silent, dominant speakers anchor groups on wrong answers, and norm violations go unsanctioned. We present Agent Sociodrama, a framework for modeling and validating evolving social dynamics in LLM agent teams. Under fixed personas, agents' emotional states and directed, asymmetric relationships are updated every round by a unified external LLM judge and shape their subsequent interactions. Across six LLM backbones and five social scenarios, we assess social fidelity through classical social-theory patterns and agreement with human judgments. The resulting trajectories exhibit structured social patterns and model-specific differences. We then introduce Trajectory-Guided Persona Recasting (TPR), which diagnoses group-dynamic failures from interaction transcripts and social trajectories and recasts agent personas before rerunning the task. Across four collaborative benchmarks, TPR improves performance from both heterogeneous and randomly initialized agent persona compositions, yielding up to 83.3%p of accuracy improvement over the initial casting. On representative distributed-information cooperation and mixed-motive bargaining tasks, we outperform SOTOPIA and AgentVerse in three of four comparisons, by up to 37.5%p in Pareto rate. These results show that validated social trajectories provide actionable signals for repairing LLM agent teams.

PDF

Open source PDF →Open local PDF →

Trouble viewing inline? Open PDF directly →

Full Text

74,239 characters extracted from source content.

Expand or collapse full text

Agent Sociodrama: Social-Dynamics-Guided Persona Recasting for LLM Agent Teams Yirui Qi Affiliation: Beihang University Affiliation: Nanyang Technological University, Singapore Xiaoming Zhang Jaehong Yoon Email: qyr888,yolixs@buaa.edu.cnjaehong.yoon@ntu.edu.sg Project Page:˜ https://lululu896.github.io/Agent-Sociodrama Abstract Large language model (LLM) agents are increasingly deployed as teams, yet they fail in characteristically social ways: information holders stay silent, dominant speakers anchor groups on wrong answers, and norm violations go unsanctioned. We present Agent Sociodrama, a framework for modeling and validating evolving social dynamics in LLM agent teams. Under fixed personas, agents’ emotional states and directed, asymmetric relationships are updated every round by a unified external LLM judge and shape their subsequent interactions. Across six LLM backbones and five social scenarios, we assess social fidelity through classical social-theory patterns and agreement with human judgments. The resulting trajectories exhibit structured social patterns and model-specific differences. We then introduce Trajectory-Guided Persona Recasting (TPR), which diagnoses group-dynamic failures from interaction transcripts and social trajectories and recasts agent personas before rerunning the task. Across four collaborative benchmarks, TPR improves performance from both heterogeneous and randomly initialized agent persona compositions, yielding up to 83.3%p of accuracy improvement over the initial casting. On representative distributed-information cooperation and mixed-motive bargaining tasks, we outperform SOTOPIA and AgentVerse in three of four comparisons, by up to 37.5%p in Pareto rate. These results show that validated social trajectories provide actionable signals for repairing LLM agent teams. 22footnotetext: Corresponding authors. 1 Introduction Large language model (LLM) agents are increasingly deployed in teams (Wu et al., 2024; Hong et al., 2024): they debate answers (Du et al., 2024), pool distributed evidence, teach one another, negotiate, and manage shared resources. Yet sustained multi-agent interaction requires more than assigning each agent an initial role. For a team to function as a social system over time, its members must form, maintain, and revise relationships in ways that remain coherent with the interaction. Existing agent systems often encode social context through static role instructions or unstructured memory, without an explicit and validated account of how directed relationships evolve over time. We take a different route. We argue that longitudinal role play requires treating a team not as a collection of fixed personas, but as a dynamic social system. A persona specifies who an agent is, but not how its relationships with others develop over time, including whom it comes to trust, avoid, align with, or oppose. These relationships must therefore be represented as evolving interaction states. The key challenge is not merely to model changing relations, but to ensure that those changes remain socially faithful over time. Figure 1: Motivation and overview of Agent Sociodrama. Agent Sociodrama models multi-agent interaction through stable personas, evolving emotions, and directed relationships. We assess the social fidelity of these dynamics against social-theory patterns and human judgments, and use social trajectories to guide team composition optimization for better collaboration. To instantiate this idea, we propose Agent Sociodrama, a longitudinal role-playing framework for LLM agent teams. Agent Sociodrama operationalizes stable personas together with evolving emotional and relational states, providing a structured account of how a team’s social configuration develops over repeated interaction. These dynamics also matter for task-oriented teams. A growing body of evidence suggests that LLM teams often fail in ways that resemble classic small-group pathologies rather than limits of individual model capability: groups that collectively possess the information required for a correct decision can still converge on an incorrect one (Pappu et al., 2026; Li et al., 2026), while individually capable agents can collectively deplete a shared resource until cooperation collapses (Piatti et al., 2024). These failures have a distinctly social character. Over repeated interaction, overlooked members can withdraw, early deference to dominant voices can harden into consensus, and unchallenged violations can progressively erode cooperation. Yet dominant approaches to improving multi-agent systems optimize prompts, topologies, or workflows (Zhou et al., 2026; Zhang et al., 2025), treating the team as a black box and leaving the social mechanisms behind these failures largely unexamined. Building on Agent Sociodrama’s explicit social trajectories, we further introduce Trajectory-Guided Persona Recasting (TPR), a casting controller that diagnoses team composition and recasts personas to seek more effective task-oriented collaboration. Our empirical study has two parts. First, we assess the social fidelity of Agent Sociodrama across six LLM drivers and five social scenarios, using a shared external LLM judge with human cross-validation. We test whether the resulting trajectories recover four established regularities of social interaction: structural balance (Heider, 1946; Cartwright & Harary, 1956), reciprocity (Gouldner, 1960), conflict-driven polarization (Krackhardt & Stern, 1988), and divergent repair-versus-escalation dynamics (Gottman, 1994; Pruitt & Kim, 2004). These analyses establish Agent Sociodrama as a socially faithful framework for longitudinal multi-agent role play. Second, across four collaborative task families, our controller converts evolving social states into persona recasting decisions that substantially improve team performance. These results demonstrate that socially faithful trajectory modeling can directly inform persona recasting and improve multi-agent collaboration. Our contributions are: • A dynamic multi-agent role-playing framework for LLM teams. Agent Sociodrama jointly models stable personas, evolving emotional states, and directed, asymmetric relational states, allowing multi-agent interaction to unfold as a sustained social process. • Cross-model validation of social fidelity. Across six LLMs and five social scenarios, we validate Agent Sociodrama against social theory and human evaluation, showing that its long-horizon relational dynamics are structured, interpretable, and socially coherent. • Trajectory-guided team composition optimization. We use social trajectories to guide persona casting, making team chemistry directly controllable. Across four collaborative task families, our controller substantially improves team performance. 2 Related Work Social simulation and coordination dynamics. LLM social simulation spans memory-based societies (Park et al., 2023), social-intelligence evaluation (Zhou et al., 2024), coordinated role play (Wang et al., 2025), and social reasoning and longitudinal interaction (Mou et al., 2025; Goel & Zhu, 2025). LLM teams exhibit information-pooling failures (Stasser & Titus, 1985), underused expertise (Pappu et al., 2026), unshared evidence (Li et al., 2026), and coordination errors (Cemri et al., 2025). BenchForm examines conformity (Weng et al., 2025), while MultiAgentBench evaluates collaboration alongside task completion (Zhu et al., 2025). Drawing on affect (Russell, 2003), reciprocity (Gouldner, 1960), group boundaries (Krackhardt & Stern, 1988), and conflict dynamics (Pruitt & Kim, 2004), Agent Sociodrama models evolving affect and directed relationships as persona-conditioned states. Their trajectories complement transcripts in diagnosing team dynamics and guiding persona recasting. Persona composition and multi-agent optimization. Personality research motivates stable behavioral specifications (McCrae & Costa, 1987), while persona composition effects depend on task structure (Keluskar et al., 2026). MetaGPT (Hong et al., 2024) and AutoGen (Wu et al., 2024) provide role-based coordination, whereas EvoAgent evolves agent configurations (Yuan et al., 2025). Automated design optimizes computational graphs (Zhuge et al., 2024), prompts and topologies (Zhou et al., 2026), workflows (Zhang et al., 2025), and code-defined agents (Hu et al., 2025); learned orchestration adapts agent activation (Dang et al., 2025). Agent Sociodrama searches persona assignments from a fixed library between task runs. Social trajectories guide proposed changes, while task outcomes determine configuration selection, linking persona composition to each team’s observed interaction dynamics. 3 Agent Sociodrama: Simulation with Measurable Social State We study multi-agent social simulation with dynamically evolving relationships. Given N agents =A1,…,ANA=\A_1,…,A_N\, each with a fixed persona, they interact over T rounds within a shared narrative scenario. While prior work on multi-agent collaboration or debate either assumes static inter-agent relationships (Li et al., 2023; Hong et al., 2024) or does not explicitly model them (Du et al., 2024; Wang et al., 2024), we let every interaction update the agents’ relationships and emotional states, which in turn shape how they communicate in subsequent rounds. This feedback process gives rise to unscripted social dynamics, such as shifting alliances, declining trust, and growing disagreement. The simulation takes as input a persona for each agent (P), initial emotional states (S0S^0), pairwise relationships (R0R^0), and a sequence of scenario events. The persona captures the stable core of an agent, its temperament, values, and characteristic style of interaction, and remains fixed throughout a run, whereas the agent’s actions, emotional states, and directed relationships evolve continually across rounds. We collect these evolving quantities into an interaction trajectory Y=(at,St,Rt)t=1T,Y= (a^t,\;S^t,\;R^t )_t=1^T, where at=aiti=1Na^t=\a_i^t\_i=1^N and St=Siti=1NS^t=\S_i^t\_i=1^N collect the actions and emotional states of all agents, and Rt=Rijti≠jR^t=\R_ij^t\_i≠ j collects all directed relationships at round t. In each round, an agent updates its emotions and relationships based on visible events, guided by its persona. These updated states and the visible history then drive its next actions. 3.1 Persona-Conditioned Agent State At round t, each agent i∈1,…,Ni∈\1,…,N\ is represented by the state (Pi,Sit,Ri⋅t)(P_i,S_i^t,R_i·^t), where PiP_i is a time-invariant persona, SitS_i^t is a dynamic emotional state, and Ri⋅t=Rijtj≠iR_i·^t=\R_ij^t\_j≠ i collects its directed relationships toward all other agents; the dot denotes all valid values of the second index j. We omit the superscript t when the round is clear from context. Time-invariant persona. Assigned before an episode and fixed throughout it, each persona Pi=(identityi,traitsi,boundaryi)P_i=(identity_i,traits_i,boundary_i) specifies the agent’s name and social role, behavioral tendencies, and behavioral constraints, respectively. The full persona guides action generation, while its traits also condition emotional and relational updates. Dynamic emotional state. Sit∈[0,1]S_i^t∈[0,1] represents the current affective valence of agent i, where 0.50.5 represents a neutral state, with lower and higher values indicating negative and positive affect, respectively. It evolves from the agent’s prior state and the events visible to it: Sit+1=US(Pi,Sit,Eit),S_i^t+1=U_S (P_i,S_i^t,E_i^t ), (1) where EitE_i^t denotes the events and interactions visible to agent i in round t, as will be detailed in Sec. 3.2. Persona conditioning allows the same event to elicit different emotional responses across agents, while incorporating SitS_i^t preserves continuity over time. Directed inter-agent relationship. Relationships are directed and need not be reciprocal: one agent may trust another even when the latter remains suspicious of the former. We represent agent i’s current relational stance toward agent j at round t as: Rijt=(favijt,statusijt,Mijt)R_ij^t= (fav_ij^t,\ status_ij^t,\ M_ij^t ). The favorability score favijt∈[−1,1]fav_ij^t∈[-1,1] tracks gradual changes in agent i’s attitude toward agent j. The categorical status records one of four discrete relationship categories, neutral, ally, rival, or enemy. The memory MijtM_ij^t retains salient interaction events that ground subsequent relational updates. For each ordered pair (i,j)(i,j), let Eijt⊆EitE_ij^t E_i^t denote the events relevant to their relationship: direct interactions between i and j, whether public or private, and other public actions by j observed by i. The relationship is then updated as Rijt+1=UR(Pi,Rijt,Eijt).R_ij^t+1=U_R (P_i,R_ij^t,E_ij^t ). (2) The concrete instantiation of USU_S and URU_R within the per-round loop is described in Sec. 3.3. 3.2 Events, Actions, and Visibility A scenario consists of ordered scenario events, ℰ=(e1,…,eT)E=(e^1,…,e^T). Each round t introduces ete^t as new context, to which agents respond while interacting with one another. Public and private interactions give agents different information, and each responds based on the events and actions visible to it. Agent AiA_i produces ai=(τi,vi,ui),τi∈(∖Ai)∪all,none,a_i= ( _i,v_i,u_i ), _i∈(A \A_i\)∪\ all, none\, (3) where uiu_i is the utterance, τi _i its recipient, and vi∈public,private,nonev_i∈\ public, private, none\ its visibility. The targets AjA_j, all, and none denote an individual recipient, the group, and silence. A public utterance is visible to all agents even when addressed to AjA_j; a private utterance is visible only to its speaker and designated recipient. Valid combinations require τi=none _i= none if and only if vi=nonev_i= none, and τi=all _i= all implies vi=publicv_i= public. The observer set is V⁡(ai)=,vi=public,Ai,Aj,τi=Aj,andvi=private∅,vi=none.V(a_i)= casesA,&v_i= public,\\ \A_i,A_j\,& _i=A_j,\ and\ v_i= private\\ ,&v_i= none. cases (4) With the scenario event e visible to all, agent AiA_i observes Ei=e∪ak|Ai∈V⁡(ak).E_i=\e\\;∪\; \a_k\; |\;A_i∈ V(a_k) \. (5) Emotional and relational dynamics can support coordination or amplify conflict without being visible in public behavior. The resulting S and R trajectories allow the casting controller (Sec. 4) to assess these dynamics and identify more effective agent compositions. 3.3 Per-Round Interaction and State Updates Each round proceeds in three stages (Fig. 2): context assembly, where code gathers the history visible to each agent; two-phase action generation, where agents first decide whether and how to speak and then speak in turn; and state update, where each agent’s emotional state and relationships are revised from what it observed. The remainder of this section details each stage. Figure 2: The Agent Sociodrama simulation with evolving emotional and relational states. Context assembly. Each agent is prompted with its local state and interaction history, augmented by the information available at the current point in the episode. The local state consists of its persona PiP_i, emotional state SitS_i^t, and directed relationships Ri⋅tR_i·^t with their associated events. The remaining context comprises globally visible scenario events and prior public utterances, actions already taken in the current round, and private messages visible only to the agents involved. Two-phase action generation. A fixed or random speaking order of agents would favor later speakers, as they observe more context before responding. We therefore decouple whether and how to speak from what to say. In Phase 1, every agent independently receives its state (P,S,R)(P,S,R) and the shared context, and output a communicative urgency u∈[0,10]u∈[0,10], a mode (public, private, or silence), a target agent, and a one-sentence communication intent. Because no agent has spoken yet in the current round, these decisions are made under identical information, preventing earlier utterances from influencing the public/private choice. In Phase 2, agents speak sequentially in descending order of urgency; silent agents are skipped and incur no inference cost, and each speaker observes the utterances of earlier speakers together with any private messages addressed to that speaker. Urgency-based ordering avoids advantages arising merely from speaking position, while allowing agents to remain silent avoids unnecessary speech in every round. Emotional update (USU_S). USU_S is a persona-conditioned LLM call that receives the trait block of PiP_i, the current state SitS_i^t, and the visible events EitE_i^t. The call returns a signed change δS _S, which is clipped to [−0.3,0.3][-0.3,0.3] to limit abrupt changes within a single round. The emotional state is then updated as Sit+1=clip⁡(Sit+δS,0,1)S_i^t+1=clip(S_i^t+ _S,0,1). Relational update (URU_R). An event in EijtE_ij^t triggers URU_R, an LLM call conditioned on the trait block of PiP_i, the current relationship RijtR_ij^t, and the triggering event. It returns a favorability change δ∈[−0.3,0.3] _fav∈[-0.3,0.3], an optional status update, a one-sentence event summary, and an importance score imp∈[1,10]imp∈[1,10]. We apply ←clip⁡(+δ,−1,1)fav (fav+ _fav,-1,1). Status combines gradual threshold-based transitions with consequential-event updates from URU_R (e.g., betrayal or alliance); when both apply, the latter takes precedence (Appendix B). Event summaries and importance scores are stored in MijM_ij with their rounds. Memory stores up to K entries as a compact relationship history for future updates without retaining the full transcript. When full, it evicts entries with the lowest retention scores, gradually removing routine interactions while preserving consequential events longer. Emotional and relational updates are computed independently within each round from visible events and persona, avoiding circular dependencies and allowing their roles to be analyzed separately. 4 The Casting Controller Given a persona composition, the simulation framework in Sec. 3 generates an interaction trajectory Y that records the agents’ actions and the evolution of their emotional states S and directed relationships R. Crucially, persona composition is itself a consequential design choice. With the same backbone and the same private information, changing only the persona casting raises GPT-5-mini’s top-1 accuracy on the Student-council hidden-profile task from 16.7% to 100.0% (Sec. 5.3). We therefore introduce a casting controller that treats persona composition as its decision variable and optimizes it guided by the measured social trajectories (Fig. 3). A casting c=(P1,…,PN)c=(P_1,…,P_N) assigns one persona to each of N agents, with each PiP_i drawn from a library of five archetypes, =balanced,dominant,analytical,agreeable,reservedP=\ balanced, dominant, analytical, agreeable, reserved\. Each task poses a distinct coordination problem: Student-council requires agents to surface distributed evidence, whereas Solar-farm negotiation requires them to reconcile competing interests. From an initial casting, the controller evaluates task performance, diagnoses the interaction trajectory, and selectively recasts agents to improve the team’s ability to solve the task. Figure 3: Trajectory-Guided Persona Recasting. Social trajectories localize team failures to guide per-agent persona recasting. Objective. Each task defines a scalar utility u⁡(c)u(c) computed from the interaction outcome. The controller aims to identify a casting that maximizes expected utility, c⋆=arg⁡maxc⁡[u⁡(c)]c = _cE[u(c)]. We estimate the expectation over repeated episodes. The utility is defined by the task’s primary metric: top-1 accuracy for the Student-council hidden-profile task, the negative L1L_1 distance to the expert ranking for NASA Moon Survival and Lost at Sea, and Pareto-efficient settlement rate for Solar-farm negotiation. Across tasks, only the task context and utility definition vary; the recasting procedure, archetype library, and diagnostic logic remain fixed. Iterative persona recasting. Searching over all ||N|P|^N possible castings is unnecessary and computationally costly. The measured interaction trajectories provide signals about which agents to recast and how. Given a initial casting c0c_0, each iteration proceeds in three steps (Algorithm 1): 1. Evaluate. Run the scenario r times under ckc_k, recording the utility u, the transcript, and the trajectories of emotional states S and directed relationships R for each run. 2. Diagnose. An external LLM examines each run’s persona composition, transcript, final outcome, and the trajectories of emotional states (St)t=1T(S^t)_t=1^T and directed relationships (Rt)t=1T(R^t)_t=1^T. It identifies whose information was suppressed, which agents dominated unhelpfully, and how the team should be recast, providing a one-sentence rationale for each change. 3. Recast. Obtain ck+1c_k+1 by majority vote over each agent’s recasting decisions across the r runs. The loop terminates when the task target is reached, no persona change is proposed, or after at most three recasts. Role of social trajectories in diagnosis. The transcript records what agents say, while the trajectories summarize how their states and relationships evolve. The controller examines these signals at three levels. At the individual level, the emotional state S helps distinguish neutral silence from withdrawal after repeated dismissal. At the dyadic level, a directed relationship RijR_ij tracks how agent i’s stance toward agent j evolves, revealing when an information holder is gradually disregarded despite outwardly polite dialogue. At the group level, patterns in R reveal emerging alliances, polarization, and shared responses to norm violations that are not explicitly articulated in the conversation. Using these signals and the transcript, the controller localizes failures to particular agents and relationships rather than falling back on generic advice such as “encourage more discussion.” 5 Experiments 5.1 Experimental Design We assess Agent Sociodrama at two levels. First, we examine whether the emotional and relational components of the interaction trajectory Y capture socially faithful dynamics across LLM backbones and interaction settings (Sec. 5.2). Second, we test whether these trajectories can guide persona recasting to improve task-oriented collaboration (Sec. 5.3). Table 1: Social dynamics fidelity in Agent Sociodrama. (a) Scenario averages over six models with five episodes each. (b) Comparisons across models used to instantiate multiple agents in Tavern scenario. Filled circles indicate conflict intensity. Full results with std. dev. are in the appendix. (a) Scenario-level social dynamics in Agent Sociodrama Scenario Conflict Intensity T Fav^T T S^T Recip. Corr. E-I index Balance (%) Repair (%) Hospital ethics review ∙∘ 0.75 0.11 0.35 −0.10-0.10 100.0 58.5 Tavern inheritance dispute ∙∘ 0.55 0.30 0.74 −0.55-0.55 95.7 51.5 Adventurer-party quest ∙∘ 0.49 0.25 0.78 −0.63-0.63 78.7 39.2 Solar-farm negotiation ∙∘ 0.43 0.15 0.91 −0.79-0.79 78.7 18.1 Demolition standoff ∙ 0.23 0.23 0.68 −2.25-2.25 71.7 18.9 (b) Tavern scenario with various models Models Conflict Intensity T Fav^T T S^T Recip. Corr. E-I index Balance (%) Repair (%) GPT-5-mini ∙∘ 0.75 0.23 0.67 −0.22-0.22 100.0 60.0 Claude Sonnet 4.6 ∙∘ 0.57 0.42 0.93 −0.53-0.53 100.0 80.0 DeepSeek V4-Flash ∙∘ 0.57 0.32 0.64 −0.52-0.52 94.0 53.3 Gemini 3.1 FL ∙∘ 0.51 0.13 0.69 −0.67-0.67 80.0 48.8 Grok 4.20 ∙∘ 0.47 0.24 0.61 −0.63-0.63 100.0 56.7 GPT-4o ∙∘ 0.45 0.46 0.92 −0.72-0.72 100.0 10.0 FavTFav^T Final favorability over directed edges STS^T Final affective valence over agents Recip. Corr. Pearson correlation of opposite-direction favorability E-I index Contrast in mean favorability across and within factions Balance % of balanced triads Repair % of deteriorated relationships recovered by the final round Detailed definitions of metrics are in Sec. D. Social dynamics fidelity evaluation. We evaluate how faithfully LLMs model emotional and relational dynamics across five scenarios spanning shared-goal coordination to open factional conflict. The scenarios draw on established paradigms in social psychology (Heider, 1946; Cartwright & Harary, 1956; Gouldner, 1960), group decision-making (Hill, 1982; Laughlin & Ellis, 1986), and negotiation research (Raiffa, 1982; Pruitt & Kim, 2004), covering consensus under divergent risk preferences, contested resources, value conflict, mixed-motive bargaining, and factional opposition. They include a hospital ethics review, inheritance dispute, adventurer-party quest, solar-farm negotiation, and demolition standoff. We evaluate six LLM backbones: GPT-4o, GPT-5-mini, Claude Sonnet 4.6, Gemini 3.1 Flash-Lite, Grok 4.20, and DeepSeek V4-Flash, with all five agents in each run instantiated by the same backbone. A GPT-4o judge evaluates the emotional and relational trajectories from every run. Each scenario–model condition comprises five independent nine-round episodes, yielding 150 team trajectories and 1,350 team-level interaction rounds. Figure 4: Team-level favorability trajectories across nine rounds. Mean favorability over all directed edges (± one standard deviation over five episodes). Relationships warm steadily under low social pressure and flatten or decline as conflict intensifies. Persona recasting evaluation. We evaluate Trajectory-Guided Persona Recasting (TPR) (Sec. 4) on four collaborative benchmarks: Student Council, NASA Moon Survival, Lost at Sea, and Solar-farm Negotiation (Tab. 5). The first three require teams to integrate privately distributed information, whereas Solar-farm Negotiation requires agents with competing preferences to reach a feasible and Pareto-efficient settlement. Each task is played by four agents (Avery, Blake, Casey, and Drew) whose private knowledge remains fixed under recasting.In the three information-integration tasks, agents rank the options individually, discuss for four rounds, and rank again; the team answer aggregates the final individual rankings with a Borda count. In Solar-farm Negotiation, agents with private scoring tables and acceptance thresholds negotiate five issues, each with three options, over four rounds. The Pareto frontier over all 35=2433^5=243 possible agreements is computed exactly by enumeration. All task metrics are computed by rule from the final outputs rather than by an LLM judge. We consider two initializations: Hetero, a manually distributed, distinct persona set, and Random, randomly sampled from the persona library. Initial casts are shared across backbones and methods (Tab. 2), and each casting is evaluated over 12 episodes using the same six LLMs as in the fidelity study. Table 2: Initial castings for Avery / Blake / Casey / Drew, shared by all backbones and compared methods. Hetero assigns four distinct archetypes, one per agent; Random is sampled from the library and may repeat archetypes. Task Hetero Random Student Council reserved / dominant / analytical / agreeable reserved / analytical / dominant / agreeable NASA Moon Survival reserved / agreeable / analytical / dominant agreeable / analytical / dominant / agreeable Lost at Sea reserved / agreeable / analytical / dominant analytical / balanced / dominant / analytical Solar-farm Negotiation dominant / agreeable / reserved / analytical reserved / agreeable / agreeable / balanced As an ablation, we evaluate a variant of Agent Sociodrama with frozen role cards, a fixed speaking order in public discussions, and no updates to emotional or relational states. We further compare ours with SOTOPIA (Zhou et al., 2024) and AgentVerse (Chen et al., 2024) on NASA Moon Survival and Solar-farm Negotiation. These two tasks capture distributed-information cooperation and mixed-motive bargaining, respectively. 5.2 Social Dynamics Fidelity on Agent Sociodrama We evaluate social dynamics fidelity from three perspectives: (1) whether the trajectories respond coherently as social pressure varies, (2) whether LLMs exhibit distinct social dynamics under the same scenario, and (3) whether their emotional and relational changes agree with human judgments. Agent Sociodrama captures consistent social structure across LLM backbones. Averaged over six LLM backbones, Tab. 1(a) shows this overall pattern. From Hospital to Demolition, mean terminal favorability drops from 0.75 to 0.23, the E-I index shifts from −0.10-0.10 to −2.25-2.25, structural balance declines from 100.0% to 71.7%, and repair falls from 58.5% to 18.9%. Balance captures local relational consistency in triads, including mutual friendship and an allied pair opposed to a third agent (Cartwright & Harary, 1956). The lower Repair rate indicates that fewer relationships that become rival or enemy recover to neutral or ally by the final round. These coordinated patterns align the simulated relationship dynamics with the social structure of each scenario, and they emerge gradually over rounds (Fig. 4). Final affective valence STS^T reflects both initial emotional state and subsequent interactions. Figure 5: Model variation across diverse social-pressure settings. Points show mean final favorability, with error bars denoting one standard deviation across repeated episodes. Reciprocity provides a dyadic view of how closely the two directions of a relationship align (Kluger et al., 2021; Xie et al., 2024). Final reciprocity rises from 0.35 in Hospital to 0.74 in Tavern, 0.78 in Quest, and 0.91 in Negotiation, and remains high at 0.68 in Demolition. Despite declining favorability, this suggests that more conflictual settings can exhibit stronger bilateral alignment. Beyond scenario-level regularities, LLMs exhibit substantial model-specific variation within the same social setting. Final favorability differs significantly across the six models in every scenario (Kruskal–Wallis tests, all p<0.01p<0.01; Fig. 5). To characterize these differences, we focus on Tavern, where outcomes are less affected by ceiling or floor effects. As shown in Tab. 1(b), GPT-5-mini maintains broadly positive and weakly polarized relationships (FavT=0.75Fav^T=0.75, E-I=−0.22=-0.22), while Claude is distinguished by high reciprocity and strong repair after deterioration (Recip.=0.93=0.93, Repair=80.0%=80.0\%). GPT-4o reveals that positive affect and high reciprocity do not necessarily imply favorable or repairable relationships, while the contrast between Gemini/Grok and DeepSeek shows that models differ in how strongly conflict amplifies polarization. These patterns suggest that model-specific social dynamics are inherently multidimensional and cannot be reduced to any single relational metric. Agreement between humans and LLMs We compare the directions of emotional and relational changes predicted by the LLM judge with annotations from thirteen human raters, who independently label each update as increasing, unchanged, or decreasing. We first note that social-dynamics judgments are inherently ambiguous: inter-human agreement is low, with Fleiss’ κ=0.288κ=0.288 for relationships and 0.1840.184 for emotions. For items where at least seven of the thirteen human raters assigned the same direction label, the LLM judge agrees with that majority label on 76.2% of relationship updates and 88.2% of emotional updates. Further, when the judge is included as an additional rater, Fleiss’ κ remains nearly unchanged, indicating minimal impact on overall agreement. 5.3 Trajectory-Guided Team Composition Optimization The fidelity analysis above establishes that Agent Sociodrama captures structured and socially meaningful dynamics across scenarios and LLM backbones. We next ask whether these trajectories can serve as actionable signals for persona recasting and improved collaboration. Tab. 3 reports results from two initial team compositions: Hetero, a composition of manually selected distinct personas, and Random, a composition randomly sampled from the same persona library. Starting from each initialization, the controller performs up to three recasts using the task-specific primary metric. Averaged over six LLM backbones, the proposed trajectory-guided persona recasting (TPR) consistently improves performance under both initializations across all four tasks, showing that the controller adapts effectively to different starting team compositions. Under Hetero, GPT-5-mini improves from 16.7% to 100.0% on Student Council, Gemini reduces its NASA Moon Survival error from 17.17 to 12.00, and GPT-5-mini reduces its Lost at Sea error from 38.17 to 28.83. Table 3: Results of Trajectory-Guided Persona Recasting (TPR) from heterogeneous and random initializations. Initial and Final report the average performance before search and after termination, respectively. Bold marks the improved endpoint within each pair; model rows average 12 episodes. Student Council Negotiation NASA Moon Lost at Sea Top-1 acc. (%) ↑ Pareto rate (%) ↑ L1L_1 ↓ L1L_1 ↓ Hetero Random Hetero Random Hetero Random Hetero Random Model Init. Final Init. Final Init. Final Init. Final Init. Final Init. Final Init. Final Init. Final TPR (Ours) w. GPT-4o 0.0 0.0 0.0 25.0 75.0 50.0 58.3 50.0 14.33 14.83 15.33 13.67 17.67 13.33 18.83 14.67 TPR (Ours) w. GPT-5-mini 16.7 100.0 41.7 100.0 0.0 8.3 0.0 8.3 19.00 18.50 21.17 20.83 38.17 28.83 35.00 27.50 TPR (Ours) w. Claude Sonnet 4.6 0.0 50.0 25.0 50.0 33.3 41.7 33.3 25.0 8.00 8.00 10.33 10.33 20.17 16.00 15.67 14.83 TPR (Ours) w. Gemini 3.1 FL 0.0 41.7 8.3 75.0 33.3 58.3 41.7 66.7 17.17 12.00 21.17 11.67 25.50 18.17 25.00 17.33 TPR (Ours) w. Grok 4.20 0.0 50.0 0.0 8.3 25.0 33.3 25.0 25.0 22.33 16.00 18.50 16.83 21.50 14.33 25.50 23.50 TPR (Ours) w. DeepSeek V4-Flash 83.3 83.3 91.7 91.7 66.7 100.0 83.3 91.7 2.83 2.83 4.17 4.17 19.67 12.67 21.83 11.33 TPR (model avg.) 16.7 54.2 27.8 58.3 38.9 48.6 40.3 44.4 13.94 12.03 15.11 12.92 23.78 17.22 23.64 18.19 w/o Social Dynamics (model avg.) 37.5 45.8 38.9 47.2 29.2 27.8 26.4 36.1 13.44 12.72 14.22 14.61 23.97 21.00 23.67 21.00 Figure 6: Performance across recasting steps. Means over six backbones and two initializations; stopped runs retain their last measured scores. Comparison with existing collaborative baselines. We next benchmark Agent Sociodrama against established collaborative-agent frameworks on downstream collaboration performance. We compare with SOTOPIA and AgentVerse using the same task-specific recasting procedure and initial casts. As shown in Tab. 4, our framework achieves the highest final mean in three of four settings, reaching Pareto rates of 48.6% and 44.4% under Hetero and Random on Negotiation. AgentVerse narrowly leads only on Random NASA (12.8912.89 vs. 12.9212.92). Social dynamics provide actionable recasting signals. In our framework, persona recasting is guided by the emotional and relational trajectories accumulated over interaction. To assess the contribution of these social dynamics, we remove the corresponding state updates while keeping the rest of the recasting pipeline unchanged. As shown in Tab. 3, full TPR consistently achieves stronger final model-average performance across tasks and both initializations, with particularly large gains on Negotiation, where the final Pareto rate increases from 27.8% to 48.6% under Hetero. This shows that the evolving S and R trajectories provide useful signals beyond the static role cards, helping the controller localize team failures and guide targeted persona recasting. Table 4: Comparison with collaborative baselines. Final means over six backbones. NASA Negotiation L1↓L_1 Pareto (%)↑ Het. Rand. Het. Rand. Ours 12.03 12.92 48.6 44.4 SOTOPIA 13.69 14.08 18.1 6.9 AgentVerse 14.06 12.89 22.2 40.3 Figure 7: A case study of trajectory-guided persona recasting on Lost at Sea (DeepSeek V4-Flash, Random). The diagnosis (❸) finds that Blake’s argument for using the fuel can for signaling was undervalued (❷); after the third recast (❹), Blake argues for it (❺), and mean L1L_1 falls from 21.83 to 11.33 (❻). The two earlier recasts made Casey balanced instead of dominant and Avery and Drew balanced instead of analytical, lowering L1L_1 to 18.00 and 14.33. Each L1L_1 averages 12 episodes; dialogue is paraphrased from logs. Case Study: Individualized Persona Recasting The recasting decisions are also interpretable at the individual level. In the Lost at Sea example in Fig. 7, the final recast assigns Blake a dominant persona, Casey an analytical persona, and Drew an agreeable persona, and Avery a balanced persona. The resulting team contains four distinct personas, and the mean L1L_1 decreases from 21.83 to 11.33 over the complete search. This example illustrates that the controller does not simply converge on a single preferred persona. These results establish team composition as a consequential and searchable design variable. Gains under both initializations show that trajectory-guided persona recasting can improve collaboration from different starting compositions, although the magnitude varies across tasks and backbones. The social trajectories validated in Sec. 5.2 provide actionable signals for identifying stronger team compositions for each setting. Complete per-iteration controller trajectories and per-backbone ablation and baseline results are reported in Appendix F. 6 Conclusion We present Agent Sociodrama, a new framework for modeling and validating evolving social dynamics in LLM agents. We track directed, asymmetric relationships and emotional states, validating cross-model comparability under a unified judge, agreement with classical social theory, and statistical reliability. The casting controller then turns these trajectories into targeted interventions: using relational and emotional trajectories and transcripts, it identifies problematic agents and relationships and recasts their personas, matching or surpassing the strongest manual casting or published reference in collaborative tasks. These results show that social dynamics can serve not only as measurable properties of LLM teams, but also as actionable signals for improving their composition. References Abdelnabi et al. (2024) Sahar Abdelnabi, Amr Gomaa, Sarath Sivaprasad, Lea Schönherr, and Mario Fritz. Cooperation, competition, and maliciousness: LLM-stakeholders interactive negotiation. In Advances in Neural Information Processing Systems, volume 37, 2024. Cartwright & Harary (1956) Dorwin Cartwright and Frank Harary. Structural balance: a generalization of Heider’s theory. Psychological Review, 63(5):277–293, 1956. Cemri et al. (2025) Mert Cemri, Melissa Z. Pan, Shuyi Yang, Lakshya A. Agrawal, Bhavya Chopra, Rishabh Tiwari, Kurt Keutzer, Aditya Parameswaran, Dan Klein, Kannan Ramchandran, Matei A. Zaharia, Joseph E. Gonzalez, and Ion Stoica. Why do multi-agent LLM systems fail? In Advances in Neural Information Processing Systems, volume 38, 2025. Chen et al. (2024) Weize Chen, Yusheng Su, Jingwei Zuo, Cheng Yang, Chenfei Yuan, Chi-Min Chan, Heyang Yu, Yaxi Lu, Yi-Hsin Hung, Chen Qian, Yujia Qin, Xin Cong, Ruobing Xie, Zhiyuan Liu, Maosong Sun, and Jie Zhou. AgentVerse: Facilitating multi-agent collaboration and exploring emergent behaviors. In International Conference on Learning Representations (ICLR), 2024. Dang et al. (2025) Yufan Dang, Chen Qian, Xueheng Luo, Jingru Fan, Zihao Xie, Ruijie Shi, Weize Chen, Cheng Yang, Xiaoyin Che, Ye Tian, Xuantang Xiong, Lei Han, Zhiyuan Liu, and Maosong Sun. Multi-agent collaboration via evolving orchestration. In Advances in Neural Information Processing Systems, volume 38, 2025. Du et al. (2024) Yilun Du, Shuang Li, Antonio Torralba, Joshua B. Tenenbaum, and Igor Mordatch. Improving factuality and reasoning in language models through multiagent debate. In Forty-first International Conference on Machine Learning (ICML), 2024. Goel & Zhu (2025) Hitesh Goel and Hao Zhu. LIFELONG SOTOPIA: Evaluating social intelligence of language agents over lifelong social interactions. arXiv preprint arXiv:2506.12666, 2025. Gottman (1994) John M. Gottman. Why Marriages Succeed or Fail. Simon & Schuster, 1994. Gouldner (1960) Alvin W. Gouldner. The norm of reciprocity: A preliminary statement. American Sociological Review, 25(2):161–178, 1960. Hall & Watson (1970) Jay Hall and Wilfred H. Watson. The effects of a normative intervention on group decision-making performance. Human Relations, 23(4):299–317, 1970. Heider (1946) Fritz Heider. Attitudes and cognitive organization. The Journal of Psychology, 21(1):107–112, 1946. Hill (1982) Gayle W. Hill. Group versus individual performance: Are N+1 heads better than one? Psychological Bulletin, 91(3):517–539, 1982. Hong et al. (2024) Sirui Hong, Mingchen Zhuge, Jonathan Chen, Xiawu Zheng, Yuheng Cheng, Jinlin Wang, Ceyao Zhang, Zili Wang, Steven Ka Shing Yau, Zijuan Lin, Liyang Zhou, Chenyu Ran, Lingfeng Xiao, Chenglin Wu, and Jürgen Schmidhuber. MetaGPT: Meta programming for a multi-agent collaborative framework. In International Conference on Learning Representations (ICLR), 2024. Hu et al. (2025) Shengran Hu, Cong Lu, and Jeff Clune. Automated design of agentic systems. In International Conference on Learning Representations (ICLR), 2025. Keluskar et al. (2026) Aryan Keluskar, Amrita Bhattacharjee, and Huan Liu. When does personality composition matter for multi-agent LLM teams? arXiv preprint arXiv:2606.27443, 2026. Kluger et al. (2021) Avraham N Kluger, Thomas E Malloy, Sarit Pery, Guy Itzchakov, Dotan R Castro, Liora Lipetz, Yaron Sela, Yaara Turjeman-Levi, Michal Lehmann, Malki New, et al. Dyadic listening in teams: Social relations model. Applied Psychology, 2021. Krackhardt & Stern (1988) David Krackhardt and Robert N. Stern. Informal networks and organizational crises: An experimental simulation. Social Psychology Quarterly, 51(2):123–140, 1988. Laughlin & Ellis (1986) Patrick R. Laughlin and Alan L. Ellis. Demonstrability and social combination processes on mathematical intellective tasks. Journal of Experimental Social Psychology, 22(3):177–189, 1986. Li et al. (2023) Guohao Li, Hasan Hammoud, Hani Itani, Dmitrii Khizbullin, and Bernard Ghanem. CAMEL: Communicative agents for “mind” exploration of large language model society. In Advances in Neural Information Processing Systems (NeurIPS), 2023. Li et al. (2026) Yuxuan Li, Aoi Naito, and Hirokazu Shirado. Systematic failures in collective reasoning under distributed information in multi-agent LLMs. In Proceedings of the 43rd International Conference on Machine Learning (ICML), 2026. McCrae & Costa (1987) Robert R. McCrae and Paul T. Costa. Validation of the five-factor model of personality across instruments and observers. Journal of Personality and Social Psychology, 52(1):81–90, 1987. Mou et al. (2025) Xinyi Mou, Jingcong Liang, Jiayu Lin, Xinnong Zhang, Xiawei Liu, Shiyue Yang, Rong Ye, Lei Chen, Haoyu Kuang, Xuanjing Huang, and Zhongyu Wei. AgentSense: Benchmarking social intelligence of language agents through interactive scenarios. In Proceedings of the 2025 Conference of the Nations of the Americas Chapter of the Association for Computational Linguistics: Human Language Technologies (Volume 1: Long Papers), p. 4975–5001, 2025. Pappu et al. (2026) Aneesh Pappu, Batu El, Hancheng Cao, Carmelo di Nolfo, Yanchao Sun, Meng Cao, and James Zou. Multi-agent teams hold experts back. In Proceedings of the 43rd International Conference on Machine Learning (ICML), volume 306 of PMLR, 2026. Park et al. (2023) Joon Sung Park, Joseph C. O’Brien, Carrie Jun Cai, Meredith Ringel Morris, Percy Liang, and Michael S. Bernstein. Generative agents: Interactive simulacra of human behavior. In Proceedings of the 36th Annual ACM Symposium on User Interface Software and Technology (UIST), 2023. Piatti et al. (2024) Giorgio Piatti, Zhijing Jin, Max Kleiman-Weiner, Bernhard Schölkopf, Mrinmaya Sachan, and Rada Mihalcea. Cooperate or collapse: Emergence of sustainable cooperation in a society of LLM agents. In Advances in Neural Information Processing Systems (NeurIPS), 2024. Pruitt & Kim (2004) Dean G. Pruitt and Sung Hee Kim. Social Conflict: Escalation, Stalemate, and Settlement. McGraw-Hill, 3rd edition, 2004. Raiffa (1982) Howard Raiffa. The Art and Science of Negotiation. Harvard University Press, 1982. Russell (2003) James A. Russell. Core affect and the psychological construction of emotion. Psychological Review, 110(1):145–172, 2003. Stasser & Titus (1985) Garold Stasser and William Titus. Pooling of unshared information in group decision making: Biased information sampling during discussion. Journal of Personality and Social Psychology, 48(6):1467–1478, 1985. Wang et al. (2024) Junlin Wang, Jue Wang, Ben Athiwaratkun, Ce Zhang, and James Zou. Mixture-of-agents enhances large language model capabilities. arXiv preprint arXiv:2406.04692, 2024. Wang et al. (2025) Xintao Wang, Heng Wang, Yifei Zhang, Xinfeng Yuan, Rui Xu, Jen-Tse Huang, Siyu Yuan, Haoran Guo, Jiangjie Chen, Shuchang Zhou, Wei Wang, and Yanghua Xiao. CoSER: Coordinating LLM-based persona simulation of established roles. In Proceedings of the 42nd International Conference on Machine Learning, volume 267 of Proceedings of Machine Learning Research, p. 64822–64858, 2025. Weng et al. (2025) Zhiyuan Weng, Guikun Chen, and Wenguan Wang. Do as we do, not as you think: the conformity of large language models. In International Conference on Learning Representations (ICLR), 2025. Wu et al. (2024) Qingyun Wu, Gagan Bansal, Jieyu Zhang, Yiran Wu, Beibin Li, Erkang Zhu, Li Jiang, Xiaoyun Zhang, Shaokun Zhang, Jiale Liu, Ahmed Hassan Awadallah, Ryen W. White, Doug Burger, and Chi Wang. AutoGen: Enabling next-gen LLM applications via multi-agent conversation. In First Conference on Language Modeling (COLM), 2024. Xie et al. (2024) Chengxing Xie, Canyu Chen, Feiran Jia, Ziyu Ye, Shiyang Lai, Kai Shu, Jindong Gu, Adel Bibi, Ziniu Hu, David Jurgens, James Evans, Philip H. S. Torr, Bernard Ghanem, and Guohao Li. Can large language model agents simulate human trust behavior? 37, 2024. doi: 10.52202/079017-0501. Yuan et al. (2025) Siyu Yuan, Kaitao Song, Jiangjie Chen, Xu Tan, Dongsheng Li, and Deqing Yang. EvoAgent: Towards automatic multi-agent generation via evolutionary algorithms. In Proceedings of the 2025 Conference of the Nations of the Americas Chapter of the Association for Computational Linguistics: Human Language Technologies (Volume 1: Long Papers), p. 6192–6217, 2025. Zhang et al. (2025) Jiayi Zhang, Jinyu Xiang, Zhaoyang Yu, Fengwei Teng, Xionghui Chen, Jiaqi Chen, Mingchen Zhuge, Xin Cheng, Sirui Hong, Jinlin Wang, Bingnan Zheng, Bang Liu, Yuyu Luo, and Chenglin Wu. AFlow: Automating agentic workflow generation. In International Conference on Learning Representations (ICLR), 2025. Zhou et al. (2026) Han Zhou, Xingchen Wan, Ruoxi Sun, Hamid Palangi, Shariq Iqbal, Ivan Vulić, Anna Korhonen, and Sercan Ö. Arık. Multi-agent design: Optimizing agents with better prompts and topologies. In International Conference on Learning Representations (ICLR), 2026. Zhou et al. (2024) Xuhui Zhou, Hao Zhu, Leena Mathur, Ruohong Zhang, Haofei Yu, Zhengyang Qi, Louis-Philippe Morency, Yonatan Bisk, Daniel Fried, Graham Neubig, and Maarten Sap. SOTOPIA: Interactive evaluation for social intelligence in language agents. In International Conference on Learning Representations (ICLR), 2024. Zhu et al. (2025) Kunlun Zhu, Hongyi Du, Zhaochen Hong, Xiaocheng Yang, Shuyi Guo, Zhe Wang, Zhenhailong Wang, Cheng Qian, Xiangru Tang, Heng Ji, and Jiaxuan You. MultiAgentBench: Evaluating the collaboration and competition of LLM agents. In Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), p. 8580–8622, 2025. Zhuge et al. (2024) Mingchen Zhuge, Wenyi Wang, Louis Kirsch, Francesco Faccio, Dmitrii Khizbullin, and Jürgen Schmidhuber. GPTSwarm: Language agents as optimizable graphs. In Forty-first International Conference on Machine Learning (ICML), 2024. Appendix A Limitations Several limitations point to future work. The fidelity evaluation relies on a single unified LLM judge and five designed scenarios; replicating the observed regularities with additional judges and structurally different scenarios would strengthen the external validity of the measurements. The optimization experiments use 12 episodes per casting, which remains small relative to the variance of LLM interaction; we therefore report consistent directions across tasks, backbones, and initializations rather than per-cell significance claims. Finally, casting is a coarse, episode-level intervention; the same measured trajectories could drive finer-grained repairs such as runtime mediation, targeted information routing, or norm enforcement, which we view as the natural next step for measurement-driven multi-agent optimization. Appendix B Categorical-Status Update After applying and clipping δ _fav, the system first updates the categorical status using the favorability thresholds statusijthr=ally,favij≥0.55,enemy,favij≤−0.50,rival,−0.50<favij≤−0.10,statusij,otherwise.status_ij^thr= casesally,&fav_ij≥ 0.55,\\ enemy,&fav_ij≤-0.50,\\ rival,&-0.50<fav_ij≤-0.10,\\ status_ij,&otherwise. cases The threshold rule captures gradual changes in the relationship. For a consequential event, URU_R may additionally return neutral, ally, rival, or enemy. This judgment is conditioned on agent i’s traits, its current favorability and status toward agent j, the triggering event, and the three most recent entries in MijM_ij. The prompt instructs URU_R to return a new category only when the event is significant enough to change the relationship. If it returns one, the event-conditioned category overrides statusijthrstatus_ij^thr; otherwise, the threshold-based result is retained. Appendix C Fidelity Evaluation Protocol The fidelity evaluation crosses five social scenarios with six LLM drivers, yielding 30 scenario–model conditions. Each condition includes five independent episodes, with five agents interacting over nine rounds per episode. The full design therefore comprises 150 team trajectories, 1,350 team-level interaction rounds, and 6,750 agent-round observations. Appendix D Metric Definitions All metrics in Tab. 1, except Repair, are computed separately for each episode and then averaged over the five episodes within each scenario–model condition. Repair is aggregated across episodes from edge counts, as defined below. Let favijT∈[−1,1]fav_ij^T∈[-1,1] and statusijTstatus_ij^T denote the favorability and categorical status of the directed edge i→ji\!→\!j at the final round T. FavTFav^T is the mean of favijTfav_ij^T over all N⁡(N−1)N(N-1) directed edges; STS^T is the mean affective valence over the N agents. Write σij=−1 _ij=-1 if statusijT∈rival,enemystatus_ij^T∈\rival,enemy\ and σij=+1 _ij=+1 otherwise. Balance ratio. Each dyad receives the sign sij=12(σij+σji)s_ij= 12( _ij+ _ji). A triad is balanced when it contains an even number of negative dyads (Cartwright & Harary, 1956): Bal=(N3)−1∑i,j,k[|(u,v)∈(i,j),(j,k),(i,k):suv<0|is even].Bal= N3^-1 _\i,j,k\1 [ |\(u,v)∈\(i,j),(j,k),(i,k)\:s_uv<0\ |\ is even ]. (6) Reciprocity r. r is the Pearson correlation between the two directions of the (N2) N2 dyads, r=corr⁡(favijTi<j,favjiTi<j).r=corr (\fav_ij^T\_i<j,\,\fav_ji^T\_i<j ). (7) It measures the linear association between opposite-direction terminal favorability across dyads; high correlation can occur with either positive or negative attitudes. E–I index. Each scenario defines two factions. With ILIL the mean favorability over within-faction ordered pairs and ELEL the mean over cross-faction pairs, E–I=EL−ILEL+IL.E--I= EL-ILEL+IL. (8) The index follows the external–internal form of Krackhardt & Stern (1988), applied to mean favorability rather than tie counts, so it is not confined to [−1,1][-1,1] and grows in magnitude once cross-faction favorability turns negative. When EL+ILEL+IL approaches zero within an episode, the index becomes unstable, which accounts for the large dispersion of a few cells in Tab. 11. Repair rate. For episode e, let e=(i,j):∃t≤T,statusije,t∈rival,enemyD_e=\(i,j):∃\,t≤ T,\ status_ij^e,t∈\rival,enemy\\ be the set of directed edges that deteriorate at least once, and let e=(i,j)∈e:statusije,T∈neutral,allyQ_e=\(i,j) _e:status_ij^e,T∈\neutral,ally\\ be the subset that recover by the final round. For each scenario–model condition, Repair is pooled across its E=5E=5 episodes: Repair=∑e=1E|e|∑e=1E|e|.Repair= _e=1^E|Q_e| _e=1^E|D_e|. (9) An episode with no deteriorated edges contributes zero to both sums rather than being assigned a repair rate of zero. If no edge deteriorates in any of the five episodes, Repair is undefined and reported as N/A. Thus, the denominator contains only directed edges that actually deteriorate, not all 20 directed edges and not the number of episodes. The metric is conditional on ruptures that occur during the interaction and does not assume that the agents begin without consensus. Panel (a) reports the unweighted mean of the six model-level Repair rates. Appendix E Persona-Recasting Experimental Setup Tasks. Tab. 5 summarizes the four collaborative tasks. Student Council, NASA Moon Survival, and Lost at Sea are information-integration tasks in which decisive knowledge is privately distributed across members; Solar-farm Negotiation is a mixed-motive task in which parties with private preferences must reach a feasible and Pareto-efficient agreement. Each task is played by four agents, Avery, Blake, Casey, and Drew. Private knowledge remains fixed under recasting.In the three information-integration tasks, each episode consists of an individual ranking before discussion, four discussion rounds, and a final individual ranking; the team answer aggregates the four final rankings with a Borda count. Top-1 accuracy checks whether the team’s first choice is the hidden-profile optimum, and L1L_1 is the sum of absolute rank differences from the expert ranking over the 15 items. In Solar-farm Negotiation, four parties with private scoring tables and private acceptance thresholds negotiate five issues with three options each over four rounds, and the final agreement is formed by an issue-wise majority over the parties’ final proposals (Abdelnabi et al., 2024). An agreement is feasible when every party meets its threshold, and Pareto-optimal when it is feasible and no other feasible agreement makes some party better off without making another worse off; because the outcome space contains only 35=2433^5=243 agreements, the Pareto frontier is computed exactly by enumeration. All task metrics are computed by rule from the final outputs rather than by an LLM judge. Initial castings. Each agent is assigned one of the five archetypes in P. Tab. 2 lists the two initial castings per task: Hetero assigns four distinct archetypes, whereas Random is sampled from the same library and may repeat archetypes. Within a task, the same initial casting is used for all backbones and for all compared methods. Recasting loop. Algo. 1 gives the diagnose-and-recast loop. Each evaluated casting is run for r=12r=12 episodes, and the controller performs at most three recasts, i.e., at most four evaluated castings (K=4K=4). The loop terminates early when the task target is reached or when no persona change is proposed; the Final value in Tab. 3 is the score of the last evaluated casting. The w/o Social Dynamics ablation replaces the social-dynamics engine with fixed-order public turn-taking and performs no emotional or relational state updates, while keeping the same tasks, initial castings, backbones, and recasting budget. SOTOPIA and AgentVerse are paired with the same recasting procedure and initial castings. Table 5: Task suite for trajectory-guided team composition optimization. Arrows indicate the preferred direction of each task-specific outcome metric. Citations identify the foundational task paradigm and, where applicable, the LLM benchmark source. For a final deal d, party i’s utility ui(d)u_i(d) is the sum of its private scores for the selected options across the five issues. Collective utility is the mean across the four parties, Ucoll(d)=14∑i=14ui(d)U_coll(d)= 14 _i=1^4u_i(d). If no final deal is produced, collective utility is set to zero. Task Collaborative challenge Outcome metric Student Council (Stasser & Titus, 1985; Pappu et al., 2026) Hidden-profile candidate selection; decisive evidence is privately distributed across members Top-1 acc. ↑ NASA Moon Survival (Hall & Watson, 1970; Pappu et al., 2026) 15-item survival ranking; expert knowledge is distributed across members L1L_1 error ↓ Lost at Sea (Hill, 1982; Pappu et al., 2026) 15-item survival ranking; critical survival knowledge can remain unheard L1L_1 error ↓ Solar-farm Negotiation (Raiffa, 1982) Private preferences must be reconciled into a feasible and Pareto-efficient agreement Pareto rate ↑ ; collective utility ↑ Table 6: Persona archetypes in the library P. Archetypes change only how an agent communicates; task role, private information, and scoring are unchanged. Archetype Behavioral description balanced Even-tempered and impartial; weighs every member’s input equally, neither dominating nor withdrawing. dominant Assertive and confident; takes the floor early, states conclusions firmly, and rarely backs down. analytical Logical and evidence-driven; insists that conclusions follow the available facts. agreeable Accommodating and conflict-averse; tends to follow an emerging consensus even with private doubts. reserved Quiet and reticent; speaks briefly and yields to louder members, sharing knowledge mainly when asked. Algorithm 1 The diagnose-and-recast loop 1: seed casting c0c_0, archetype library P, repetitions r, budget K 2: for k=0,1,…,K−1k=0,1,…,K-1 do 3: (u(j),transcript(j),S(j),R(j))j=1r←\(u^(j),transcript^(j),S^(j),R^(j))\_j=1^r← Engine(ckc_k) run r times ⊳ Evaluate (Sec. 3) 4: for j=1,…,rj=1,…,r do 5: ρ(j)←ρ^(j)← Diagnose(ckc_k, transcript(j), u(j)u^(j), S(j)S^(j), R(j)R^(j)) ⊳ frozen GPT-4o; recast proposal 6: end for 7: ck+1←c_k+1← MajorityVote(ρ(1),…,ρ(r)ρ^(1),…,ρ^(r)) ⊳ merged per member 8: if target met or ck+1=ckc_k+1=c_k then 9: break 10: end if 11: end for 12: return last evaluated casting ⊳ no rollback to earlier iterations Appendix F Complete Persona-Recasting Results Tab. 7 reports the complete per-iteration trajectories of TPR for every task, backbone, and initialization. Tab. 8 and Tab. 9 report the per-backbone results underlying the model averages of the ablation in Tab. 3 and of the external-framework comparison in Tab. 4. All numbers are computed from the same 12-episode runs as the main text. Table 7: Complete per-iteration trajectories of TPR. Iteration 0 is the initial casting; each subsequent column reports the mean score (12 episodes) after one recast. “–” indicates that the loop had already terminated (target reached or no persona change proposed). The last evaluated iteration of each run corresponds to the Final column of Tab. 3. Hetero Random Task Model It. 0 It. 1 It. 2 It. 3 It. 0 It. 1 It. 2 It. 3 Student Council Top-1 acc. (%) ↑ GPT-4o 0.0 0.0 0.0 0.0 0.0 25.0 0.0 25.0 GPT-5-mini 16.7 100.0 – – 41.7 100.0 – – Claude Sonnet 4.6 0.0 33.3 8.3 50.0 25.0 8.3 50.0 – Gemini 3.1 FL 0.0 16.7 33.3 41.7 8.3 75.0 – – Grok 4.20 0.0 0.0 50.0 – 0.0 0.0 16.7 8.3 DeepSeek V4-Flash 83.3 – – – 91.7 – – – Solar-farm Negotiation Pareto rate (%) ↑ GPT-4o 75.0 50.0 33.3 50.0 58.3 33.3 50.0 50.0 GPT-5-mini 0.0 0.0 25.0 8.3 0.0 8.3 0.0 8.3 Claude Sonnet 4.6 33.3 83.3 50.0 41.7 33.3 8.3 8.3 25.0 Gemini 3.1 FL 33.3 41.7 50.0 58.3 41.7 41.7 50.0 66.7 Grok 4.20 25.0 33.3 25.0 33.3 25.0 0.0 8.3 25.0 DeepSeek V4-Flash 66.7 91.7 75.0 100.0 83.3 83.3 91.7 91.7 NASA Moon Survival L1L_1 ↓ GPT-4o 14.33 15.67 14.50 14.83 15.33 13.67 – – GPT-5-mini 19.00 21.67 21.50 18.50 21.17 21.00 20.50 20.83 Claude Sonnet 4.6 8.00 – – – 10.33 – – – Gemini 3.1 FL 17.17 15.50 16.33 12.00 21.17 17.00 11.67 – Grok 4.20 22.33 16.33 20.50 16.00 18.50 20.33 18.83 16.83 DeepSeek V4-Flash 2.83 – – – 4.17 – – – Lost at Sea L1L_1 ↓ GPT-4o 17.67 13.17 11.00 13.33 18.83 13.17 13.00 14.67 GPT-5-mini 38.17 27.50 35.67 28.83 35.00 27.50 34.50 27.50 Claude Sonnet 4.6 20.17 19.67 18.17 16.00 15.67 16.17 15.83 14.83 Gemini 3.1 FL 25.50 21.17 28.83 18.17 25.00 18.50 24.83 17.33 Grok 4.20 21.50 19.00 17.83 14.33 25.50 21.33 23.50 23.50 DeepSeek V4-Flash 19.67 18.33 24.67 12.67 21.83 18.00 14.33 11.33 Table 8: Per-backbone results of the w/o Social Dynamics ablation, in the same format as Tab. 3. The model-average row equals the w/o Social Dynamics row of Tab. 3. Each cell averages 12 episodes; bold marks the better endpoint within each pair. Student Council Negotiation NASA Moon Lost at Sea Top-1 acc. (%) ↑ Pareto rate (%) ↑ L1L_1 ↓ L1L_1 ↓ Hetero Random Hetero Random Hetero Random Hetero Random Model Init. Final Init. Final Init. Final Init. Final Init. Final Init. Final Init. Final Init. Final GPT-4o 0.0 0.0 0.0 0.0 25.0 16.7 16.7 25.0 15.33 13.00 15.00 16.50 21.33 15.33 20.33 15.83 GPT-5-mini 100.0 100.0 100.0 100.0 8.3 0.0 0.0 0.0 16.00 18.00 19.83 23.67 32.33 32.50 32.83 31.50 Claude Sonnet 4.6 25.0 33.3 33.3 16.7 25.0 33.3 41.7 25.0 12.17 12.17 15.67 13.67 19.17 20.33 22.00 18.83 Gemini 3.1 FL 0.0 25.0 0.0 33.3 41.7 25.0 33.3 50.0 16.33 13.67 17.00 15.83 31.50 23.33 31.67 26.00 Grok 4.20 0.0 16.7 0.0 33.3 16.7 41.7 0.0 33.3 20.67 19.33 17.67 17.83 23.50 21.83 24.33 22.67 DeepSeek V4-Flash 100.0 100.0 100.0 100.0 58.3 50.0 66.7 83.3 0.17 0.17 0.17 0.17 16.00 12.67 10.83 11.17 Model avg. 37.5 45.8 38.9 47.2 29.2 27.8 26.4 36.1 13.44 12.72 14.22 14.61 23.97 21.00 23.67 21.00 Table 9: Per-backbone results of external role-play frameworks paired with the same persona-recasting procedure and initial castings. The model-average Final values equal those in Tab. 4. Each cell averages 12 episodes; bold marks the better endpoint within each pair. NASA Moon (L1↓L_1 ) Negotiation (Pareto rate, %↑ ) Hetero Random Hetero Random Framework Model Init. Final Init. Final Init. Final Init. Final SOTOPIA GPT-4o 15.33 16.83 16.33 15.67 0.0 0.0 0.0 0.0 GPT-5-mini 18.33 18.67 16.67 18.67 0.0 0.0 0.0 0.0 Claude Sonnet 4.6 12.33 12.33 12.00 12.00 0.0 8.3 0.0 0.0 Gemini 3.1 FL 15.67 16.50 20.33 18.33 8.3 0.0 16.7 0.0 Grok 4.20 20.67 17.83 16.83 19.67 0.0 16.7 25.0 8.3 DeepSeek V4-Flash 0.00 0.00 0.17 0.17 50.0 83.3 58.3 33.3 Model avg. 13.72 13.69 13.72 14.08 9.7 18.1 16.7 6.9 AgentVerse GPT-4o 17.00 17.00 13.83 13.83 16.7 8.3 0.0 25.0 GPT-5-mini 16.50 18.83 15.83 18.50 8.3 0.0 0.0 33.3 Claude Sonnet 4.6 12.17 12.17 16.33 12.83 0.0 0.0 8.3 0.0 Gemini 3.1 FL 20.83 18.67 21.67 17.33 58.3 41.7 50.0 50.0 Grok 4.20 19.33 16.83 19.00 13.83 50.0 58.3 16.7 33.3 DeepSeek V4-Flash 0.83 0.83 1.00 1.00 8.3 25.0 50.0 100.0 Model avg. 14.44 14.06 14.61 12.89 23.6 22.2 20.8 40.3 Ours GPT-4o 14.33 14.83 15.33 13.67 75.0 50.0 58.3 50.0 GPT-5-mini 19.00 18.50 21.17 20.83 0.0 8.3 0.0 8.3 Claude Sonnet 4.6 8.00 8.00 10.33 10.33 33.3 41.7 33.3 25.0 Gemini 3.1 FL 17.17 12.00 21.17 11.67 33.3 58.3 41.7 66.7 Grok 4.20 22.33 16.00 18.50 16.83 25.0 33.3 25.0 25.0 DeepSeek V4-Flash 2.83 2.83 4.17 4.17 66.7 100.0 83.3 91.7 Model avg. 13.94 12.03 15.11 12.92 38.9 48.6 40.3 44.4 Appendix G Human Annotation Protocol We sampled 25 dialogue excerpts by stratified sampling over scenarios, LLM drivers, the two measured dimensions (R and S), and the three change directions (increase, unchanged, decrease), and added one attention-check item with an unambiguous answer. Thirteen raters independently labeled each excerpt with the same three-way direction scale used by the judge; all raters passed the attention check. A human-consensus label is assigned by strict majority (≥7≥ 7 of 13 raters), and judge agreement is computed only on items with a consensus label. Tab. 10 summarizes the results. Table 10: Agreement between the LLM judge and the human majority label. Dimension Consensus items Judge matches Agreement Human Fleiss’ κ Relationship (R) 21 16 76.2% 0.288 Emotion (S) 17 15 88.2% 0.184 Inter-human agreement is low, which reflects the inherent ambiguity of reading relational and emotional change from short excerpts rather than a failure of the raters; on the items where humans do converge, the judge follows the majority in most cases. On the relationship dimension, the judge labels a larger share of updates as increasing than the human raters do (about 76% versus about 40%). Because every condition in Tab. 1 and Tab. 11 is scored by the same judge, such a shared directional tendency affects absolute favorability levels but not the cross-scenario ordering or the cross-model comparisons on which our conclusions rest. Appendix H Full Social-Dynamics Fidelity Results Tab. 11 reports all 30 scenario–model conditions underlying Tab. 1, as mean ± standard deviation over five complete nine-round episodes per condition. Panel (a) of Tab. 1 is the equally weighted mean of the six model rows within each scenario, and panel (b) corresponds to the Tavern rows. The last column reports the Kruskal–Wallis test of final favorability across the six models within each scenario (df =5=5). Table 11: Full social-dynamics fidelity results (mean ± s.d., n=5n=5 episodes per condition). Metric definitions follow Sec. D. KW: Kruskal–Wallis test of FavTFav^T across models within each scenario. Scenario Model FavTFav^T STS^T Recip. Corr. E-I index Balance (%) Repair (%) KW Hospital ethics review GPT-5-mini 0.82±0.030.82± 0.03 0.02±0.020.02± 0.02 0.36±0.260.36± 0.26 −0.10±0.02-0.10± 0.02 100.0±0.0100.0± 0.0 40.0±20.040.0± 20.0 Claude Sonnet 4.6 0.86±0.060.86± 0.06 0.18±0.070.18± 0.07 0.28±0.170.28± 0.17 −0.08±0.05-0.08± 0.05 100.0±0.0100.0± 0.0 73.3±38.973.3± 38.9 DeepSeek V4-Flash 0.79±0.040.79± 0.04 0.24±0.060.24± 0.06 0.49±0.210.49± 0.21 −0.17±0.04-0.17± 0.04 100.0±0.0100.0± 0.0 80.0±40.080.0± 40.0 H=23.36H=23.36 Gemini 3.1 FL 0.74±0.030.74± 0.03 0.05±0.060.05± 0.06 0.76±0.120.76± 0.12 −0.14±0.08-0.14± 0.08 100.0±0.0100.0± 0.0 58.0±15.758.0± 15.7 p=0.0003p=0.0003 Grok 4.20 0.54±0.040.54± 0.04 0.04±0.020.04± 0.02 −0.08±0.31-0.08± 0.31 0.05±0.080.05± 0.08 100.0±0.0100.0± 0.0 39.7±9.739.7± 9.7 GPT-4o 0.74±0.020.74± 0.02 0.11±0.050.11± 0.05 0.32±0.310.32± 0.31 −0.15±0.04-0.15± 0.04 100.0±0.0100.0± 0.0 60.0±49.060.0± 49.0 Tavern inheritance dispute GPT-5-mini 0.75±0.030.75± 0.03 0.23±0.050.23± 0.05 0.67±0.280.67± 0.28 −0.22±0.07-0.22± 0.07 100.0±0.0100.0± 0.0 60.0±49.060.0± 49.0 Claude Sonnet 4.6 0.57±0.070.57± 0.07 0.42±0.110.42± 0.11 0.93±0.020.93± 0.02 −0.53±0.13-0.53± 0.13 100.0±0.0100.0± 0.0 80.0±24.580.0± 24.5 DeepSeek V4-Flash 0.57±0.050.57± 0.05 0.32±0.040.32± 0.04 0.64±0.110.64± 0.11 −0.52±0.05-0.52± 0.05 94.0±12.094.0± 12.0 53.3±24.553.3± 24.5 H=19.66H=19.66 Gemini 3.1 FL 0.51±0.100.51± 0.10 0.13±0.070.13± 0.07 0.69±0.120.69± 0.12 −0.67±0.12-0.67± 0.12 80.0±16.780.0± 16.7 48.8±25.748.8± 25.7 p=0.0014p=0.0014 Grok 4.20 0.47±0.050.47± 0.05 0.24±0.100.24± 0.10 0.61±0.150.61± 0.15 −0.63±0.12-0.63± 0.12 100.0±0.0100.0± 0.0 56.7±38.956.7± 38.9 GPT-4o 0.45±0.020.45± 0.02 0.46±0.130.46± 0.13 0.92±0.020.92± 0.02 −0.72±0.06-0.72± 0.06 100.0±0.0100.0± 0.0 10.0±20.010.0± 20.0 Adventurer-party quest GPT-5-mini 0.54±0.040.54± 0.04 0.22±0.060.22± 0.06 0.77±0.140.77± 0.14 −0.55±0.11-0.55± 0.11 88.0±14.788.0± 14.7 30.5±17.230.5± 17.2 Claude Sonnet 4.6 0.61±0.040.61± 0.04 0.27±0.060.27± 0.06 0.78±0.090.78± 0.09 −0.46±0.11-0.46± 0.11 100.0±0.0100.0± 0.0 80.8±10.580.8± 10.5 DeepSeek V4-Flash 0.53±0.030.53± 0.03 0.29±0.020.29± 0.02 0.81±0.110.81± 0.11 −0.54±0.12-0.54± 0.12 80.0±16.780.0± 16.7 40.0±28.940.0± 28.9 H=22.39H=22.39 Gemini 3.1 FL 0.40±0.060.40± 0.06 0.15±0.060.15± 0.06 0.70±0.120.70± 0.12 −0.70±0.15-0.70± 0.15 60.0±15.560.0± 15.5 46.3±9.746.3± 9.7 p=0.0004p=0.0004 Grok 4.20 0.40±0.060.40± 0.06 0.37±0.080.37± 0.08 0.78±0.060.78± 0.06 −0.77±0.10-0.77± 0.10 72.0±16.072.0± 16.0 15.2±11.015.2± 11.0 GPT-4o 0.48±0.020.48± 0.02 0.23±0.040.23± 0.04 0.87±0.030.87± 0.03 −0.77±0.07-0.77± 0.07 72.0±14.772.0± 14.7 22.4±19.622.4± 19.6 Solar-farm negotiation GPT-5-mini 0.52±0.060.52± 0.06 0.19±0.050.19± 0.05 0.86±0.070.86± 0.07 −0.43±0.17-0.43± 0.17 94.0±12.094.0± 12.0 27.0±6.027.0± 6.0 Claude Sonnet 4.6 0.54±0.030.54± 0.03 0.20±0.030.20± 0.03 0.84±0.070.84± 0.07 −0.53±0.08-0.53± 0.08 88.0±14.788.0± 14.7 18.0±10.718.0± 10.7 DeepSeek V4-Flash 0.47±0.030.47± 0.03 0.12±0.020.12± 0.02 0.94±0.020.94± 0.02 −0.77±0.12-0.77± 0.12 60.0±0.060.0± 0.0 19.8±5.919.8± 5.9 H=22.56H=22.56 Gemini 3.1 FL 0.29±0.060.29± 0.06 0.11±0.040.11± 0.04 0.98±0.020.98± 0.02 −0.96±0.22-0.96± 0.22 92.0±16.092.0± 16.0 6.7±13.36.7± 13.3 p=0.0004p=0.0004 Grok 4.20 0.39±0.040.39± 0.04 0.10±0.060.10± 0.06 0.92±0.020.92± 0.02 −1.04±0.10-1.04± 0.10 72.0±14.772.0± 14.7 24.2±17.624.2± 17.6 GPT-4o 0.39±0.030.39± 0.03 0.17±0.090.17± 0.09 0.93±0.020.93± 0.02 −1.03±0.13-1.03± 0.13 66.0±4.966.0± 4.9 12.9±12.412.9± 12.4 Demolition standoff GPT-5-mini 0.28±0.040.28± 0.04 0.29±0.050.29± 0.05 0.24±0.180.24± 0.18 −2.01±0.44-2.01± 0.44 94.0±12.094.0± 12.0 26.1±3.826.1± 3.8 Claude Sonnet 4.6 0.27±0.050.27± 0.05 0.23±0.050.23± 0.05 0.80±0.040.80± 0.04 −2.44±0.56-2.44± 0.56 54.0±4.954.0± 4.9 16.8±6.416.8± 6.4 DeepSeek V4-Flash 0.28±0.030.28± 0.03 0.23±0.060.23± 0.06 0.74±0.050.74± 0.05 −2.33±0.50-2.33± 0.50 94.0±12.094.0± 12.0 41.3±7.541.3± 7.5 H=18.73H=18.73 Gemini 3.1 FL 0.08±0.030.08± 0.03 0.14±0.040.14± 0.04 0.80±0.020.80± 0.02 −1.20±18.13-1.20± 18.13 64.0±4.964.0± 4.9 11.8±6.611.8± 6.6 p=0.0022p=0.0022 Grok 4.20 0.19±0.010.19± 0.01 0.25±0.030.25± 0.03 0.71±0.020.71± 0.02 −3.43±0.37-3.43± 0.37 68.0±17.268.0± 17.2 14.7±14.614.7± 14.6 GPT-4o 0.28±0.020.28± 0.02 0.27±0.060.27± 0.06 0.78±0.080.78± 0.08 −2.08±0.28-2.08± 0.28 56.0±4.956.0± 4.9 2.9±5.72.9± 5.7 Fig. 8 visualizes the same 30 conditions metric by metric, and Fig. 9 overlays the six models within each scenario. Each radar axis is min–max normalized over all 30 conditions, and polarization is plotted as −-E-I so that larger values indicate stronger factional separation. The scenario ordering is visible as a shrinking favorability and repair profile from Hospital to Demolition, while model profiles remain distinct within every scenario. Figure 8: Scenario × model means of the six fidelity metrics (n=5n=5 episodes per cell). Figure 9: Social-dynamics profiles of the six models in each scenario. Axes are min–max normalized across all 30 conditions; polarization is −-E-I. Appendix I Dynamic Trajectories Terminal metrics hide how emotional states and relationships evolve. The figures below are computed from the same 150 episodes as Tab. 11 (five complete nine-round episodes per scenario–model condition). Fig. 10 shows per-agent emotional trajectories, Fig. 4 shows team-level mean favorability over all directed edges, and Fig. 11 reports the proportion of agent-episode final states with ST≤0.1S^T≤ 0.1. Figure 10: Emotional-state trajectories of every agent across nine rounds. Lines show the mean over five episodes and shaded regions one standard deviation; the dashed line marks S=0.1S=0.1. Figure 11: Crash rate: proportion of agent-episode final states with ST≤0.1S^T≤ 0.1 (25 agent-episodes per cell).

Instant research discovery