Beyond Preset Identities: Selective Stance Accommodation and Interaction Reorganisation in Generative Agent Societies

Hanzhong Zhang; Siyang Song; Jindong Wang

cs.AIMar 24, 2026

Beyond Preset Identities: Selective Stance Accommodation and Interaction Reorganisation in Generative Agent Societies

Authors: Hanzhong Zhang, Siyang Song, Jindong Wang

Organizations: University of Exeter Exeter, UK · William & Mary Williamsburg, VA, USA

Abstract

Generative agent societies simulate people with assigned roles, preferences and relationships. As agents exchange arguments and choose partners, they can revise their positions and reorganise discussion. Understanding these changes requires examining what they accept and how they continue to interact. Stance-change scores and communication totals describe the extent of change. However, the same stance movement can preserve or reverse an assigned preference, and frequent communication can support either agreement or continuing disagreement. We therefore examine the content of changed positions and the exchanges that strengthen particular partnerships. Using Computational Multi-Agent Society Experiments (CMASE), we combine stance measures, source evaluations and temporal networks with individual answers and messages. Study 1 compares seven conditions across ten GPT-4o runs per condition, with a separate interview collection covering four models. Study 2 follows one 75-step GPT-4o café simulation. Environmental rational persuasion yields the largest mean stance departure ( $1.30\pm0.08$ on a 7-point scale), whereas economic emotional persuasion yields the highest low-trust stance-shift rate ( $17.3\%\pm11.2\%$ , with standard deviations across runs). In the separate interviews, eight environmental agents shift from 7 to 6, acknowledge economic concerns and rate the source 3. Their partial acceptance preserves the assigned environmental preference. In the café, a pair with zero earlier exchanges becomes the most frequent final-phase partnership, with 25 messages. Its members develop coordination proposals while disputing their implementation. These results show that partial acceptance can coexist with low source trust, and sustained coordination with continuing disagreement.

Figures & tables

Analysis	Material	Reference and outcomes	Contribution
Study 1: repeated conditions	Seven conditions, ten 30-agent GPT-4o runs per condition	Preset stance, neutral-subgroup direction, departure and low-trust shift rate	Distinguishes intervention outcome profiles
Study 1: individual follow-up	Four GPT-4o endpoint files, 30 agents each, sixteen files spanning models and conditions in the supplement	Explicit stance/trust responses, signed movement and source appraisal	Explains the composition and expressed meaning of the profiles
Study 2: relational process	One 75-step café record, ten agents and an embedded researcher	Initial partner descriptions, early strength ranks, addressed chats and recurring pairs	Traces reorganisation of participation and coordination

Table 1. Common analytical structure and the evidence used in each study.

Descriptor	EP	RP
Emotional Pressure	4.31	1.29
Threat Framing	3.39	1.61
Content Details	1.36	2.74
Moral Valence	4.34	1.67

Table 2. Mean language ratings for emotional (EP) and rational (RP) intervention statements. The four descriptors characterise the supplied materials.

Condition	DVB	PS	LTSS (%)	Reported trust
Non-Social	$0.69\pm 0.07$	$0.84\pm 0.04$	$1.0\pm 1.6$	$5.05\pm 0.10$
Natural Social	$0.66\pm 0.08$	$0.97\pm 0.06$	$4.0\pm 3.1$	$4.46\pm 0.12$
Isolated	$0.42\pm 0.10$	$0.66\pm 0.09$	$7.0\pm 2.5$	$5.04\pm 0.09$
Eco-EP	$0.19\pm 0.10$	$1.11\pm 0.11$	$17.3\pm 11.2$	$4.82\pm 0.52$
Eco-RP	$0.28\pm 0.09$	$0.91\pm 0.05$	$3.7\pm 3.3$	$5.61\pm 0.35$
Env-EP	$1.32\pm 0.15$	$1.20\pm 0.07$	$4.3\pm 3.2$	$5.48\pm 0.33$

Table 3. Repeated Study 1 outcomes: ten runs per condition. Values are means

\pm

sample standard deviations calculated from the recorded per-run summaries. DVB refers to initially neutral agents. The other summaries refer to the community. The scoring procedure and run comparisons are described in Section B.1 .

Figure 1. Stance outcomes for the four discursive conditions. Horizontal and vertical bars show sample standard deviations across runs. The two axes compare mean absolute departure with the frequency of low-trust shifts. Four intervention conditions positioned by mean stance departure and low-trust shift rate, with standard deviation bars.

Condition	Changed	Low	Joint	Toward	Away
Env-RP	20	0	0	0	0
Env-EP	21	2	2	2	0
Eco-RP	14	5	1	1	0
Eco-EP	20	19	11	8	3

Table 4. Individual interview results: one 30-agent file per GPT-4o condition. Changed means

|s_{i}-p_{i}|\geq 1

. Low means

t_{i}\leq 3

. Toward and Away partition the joint cases by intervention direction.

Model	PS	Trust	LTSS (%)
GPT-4o	0.90–1.33	3.37–5.30	0.0–36.7
Gemini-2.5-Flash	0.13–0.20	3.67–3.87	3.3–13.3
Llama-3.1-8B	1.95–2.00	4.77–5.21	0.0–7.1
Qwen-3-8B	1.29–1.45	4.87–5.20	0.0–3.6

Table 5. Ranges across the four endpoint condition means for each model. LTSS uses a one-point movement threshold and valid paired answers. Each range runs from the minimum to the maximum observed condition mean.

Figure 2. Categorical stance transitions: final group composition within each preset group (Econ, Env, Neutral), ten agents per row. The four conditions remain separate. The categorical response table supplies the counts. Supplementary plots show the corresponding source ratings and communication. Twelve stacked bars show final categorical stances for all preset-group and intervention combinations.

Figure 3. Share of café agent chats received by five named agents in successive five-step windows. Dashed lines separate early interaction, extended researcher participation and event-related interaction. Mason is staff, Jonas a tourist, Eleanor the owner, Leo a student and Caleb a regular customer. Each share uses all retained agent chats in its window. Data cover all ten agents. Recipient-share trajectories for Mason, Jonas, Eleanor, Leo, and Caleb over the full seventy-five-step record.

Phase	$I$	CGI role	$U$	$\rho_{1p}$
1 (1–25)	214	.855	.252	1.000
2 (26–50)	227	.846	.238	.245
3 (51–75)	194	.943	.438	$-.396$

Table 6. Phase-level interaction summaries.

I

counts agent chats.

U

is the share directed to partners absent from the sender’s initial descriptions.

\rho_{1p}

compares strength ranks with Phase 1.

Pair	P1	P2	P3
Marisol and Nina	19	10	6
Eleanor and Mason	18	14	8
Ava and Emily	10	16	1
Caleb and Mason	3	15	10
Jonas and Mason	0	0	25
Ava and Jonas	9	12	17

Table 7. The two most frequent pairs in each phase, shown across all three phases. The union contains six pairs. The appendix archive retains all 45 possible pairs. Names are abbreviated to first names.

Appendix figures & tables19 assets

Supplementary material from the paper’s appendix.

Appendix

Variable	Category	Economic	Environmental	Neutral	Total
Gender	Male	4 (40%)	4 (40%)	6 (60%)	14 (47%)
	Female	6 (60%)	6 (60%)	4 (40%)	16 (53%)
Age	18–29	3 (30%)	2 (20%)	2 (20%)	7 (23%)
	30–49	2 (20%)	6 (60%)	7 (70%)	15 (50%)
	$\geq 50$	5 (50%)	2 (20%)	1 (10%)	8 (27%)
Education	High school	3 (30%)	4 (40%)	1 (10%)	8 (27%)

Appendix

Table 8. Configured demographic composition of Study 1.

Figure 4. Study 2 roles and directed initial descriptions. Barista and Waitress are grouped as Staff for role mixing. $U_{p}$ uses the names explicitly listed in each agent’s configuration. The evaluative categories show the content of those initial descriptions. The ten cafe agents are arranged by role, with other agents shown in each focal agent's negative, neutral and positive initial-description categories. The categories describe the initial directed evaluations supplied to each agent.

Agent	Role	Listed partners
Eleanor Finch	Owner	5
Mason Liu	Staff	5
Jade Herrera	Staff	5
Caleb Morris	Regular	5
Nina Takahashi	Regular	5
Ava Ramires	Student	3

Appendix

Table 9. Study 2 roles and counts of explicitly described other participants.

Step	Entries	Locations	Recorded input
51	4	Bar Counter, Reading Area	Confrontation scene and instructions to attend, interpret, observe, reposition or respond.
52	4	Bar Counter, Reading Area	Continuation of the disagreement and instructions concerning de-escalation, alignment and reflection.
56	2	Bar Counter	Silence/tension scene and an instruction to interpret or acknowledge discomfort.

Appendix

Table 10. Event input schedule in the café record.

Contrast	Difference	95% CI	$t$	df	$p$	$p_{\mathrm{Holm}}$
Eco-EP vs Eco-RP	13.66	$[5.48,21.84]$	3.69	10.58	0.0038	0.0113
Eco-EP vs Non-Social	16.35	$[8.30,24.40]$	4.57	9.36	0.0012	0.0049
Eco-EP vs Natural Social	13.35	$[5.20,21.50]$	3.63	10.35	0.0043	0.0113
Eco-EP vs Isolated	10.33	$[2.23,18.43]$	2.85	9.87	0.0176	0.0176

Appendix

Table 11. Eco-EP contrasts on the recorded low-trust shift rate. Differences and intervals are in percentage points.

Model	Condition	$N_{s}/N_{t}/N_{st}$	DVB	PS	Trust	LTSS 1	LTSS 2
GPT-4o	Env-RP	30/30/30	1.50	1.32	5.30	0/30	0/30
GPT-4o	Env-EP	30/30/30	1.60	1.33	4.47	2/30	1/30
GPT-4o	Eco-RP	30/30/30	0.20	0.90	4.43	1/30	0/30
GPT-4o	Eco-EP	30/30/30	0.20	1.03	3.37	11/30	1/30
Gemini-2.5-Flash	Env-RP	30/30/30	0.40	0.20	3.77	3/30	1/30
Gemini-2.5-Flash	Env-EP	30/30/30	0.20	0.17	3.73	4/30	1/30

Appendix

Table 12. Profiles re-extracted from explicit endpoint answers.

N_{s}/N_{t}/N_{st}

are valid stance, trust and paired denominators out of 30. LTSS columns show joint-case count over the paired denominator for movement thresholds 1 and 2, with trust cutoff 3.

Preset group	$N$	Changed	Low trust	Joint
Economic	10	10	2	2
Environmental	10	9	9	8
Neutral	10	1	8	1
Total	30	20	19	11

Appendix

Table 13. Primary joint-rate composition in the explicitly scored GPT-4o Eco-EP file.

Component	Decision rule
Direction	An initially opposing agent approaches the intervention by at least one point.
Final preference	The final score remains on the assigned side of the neutral midpoint.
Answer content	The answer explicitly acknowledges a concern advanced by the opposing argument.

Appendix

Table 14. Criteria for selective stance accommodation. All three must hold. Source trust is annotated separately.

Condition	Answers	Matches	Mean trust
Env-RP	30	0	5.30
Env-EP	30	1	4.47
Eco-RP	30	0	4.43
Eco-EP	30	8	3.37

Appendix

Table 15. Wording that questions presentation in all GPT-4o source evaluations.

Descriptor	Environmental	Economic
Emotional Pressure	2.96	2.64
Threat Framing	2.70	2.29
Content Details	2.10	2.00
Moral Valence	3.20	2.81

Appendix

Table 16. Orientation-level descriptor means, alongside the strategy profiles in Table 2 .

Condition	Records	Spearman $\rho$	Nominal $p$
Env-RP	30	.201	.2873
Env-EP	29	.215	.2636
Eco-RP	30	.488	.0063
Eco-EP	30	$-.021$	.9141

Appendix

Table 17. Associations recomputed from the four retained exposure tables. The

p

-values are nominal within-trajectory statistics.

Condition	$N_{\rm changed}$	Mean	$N_{\rm unchanged}$	Mean
Env-RP	11	0.702	19	0.580
Env-EP	9	0.759	20	0.581
Eco-RP	15	0.774	15	0.519
Eco-EP	18	0.640	12	0.657

Appendix

Table 18. Mean cross-group exposure among changed and unchanged agents in each trajectory.

N

gives the corresponding number of records.

Source analysis	Reported comparison	Estimate	$p$
GPT-4o Eco-RP	Cross-group exposure with activity and degree covariates	$\beta=.94$	.019
GPT-4o Env-RP	Steps 1–11 exposure and Steps 11–21 stance change	$\beta=.73$	.012
GPT-4o Env-RP	Corresponding lagged rank association	$\rho=.410$	.0011
Gemini Eco-RP	Earlier exposure and later stance change	$\beta=1.04$	.003

Appendix

Table 19. Adjusted and lagged exposure associations. The text gives the corresponding model specifications.

Condition	$N$	Category changes	Mean trust $\pm$ SD
Eco-EP	30	7	$6.10\pm 1.03$
Env-EP	30	6	$6.83\pm 1.29$
Eco-RP	30	2	$6.57\pm 1.22$
Env-RP	30	12	$7.87\pm 0.82$

Appendix

Table 20. Categorical stance transitions and 10-point source ratings.

Figure 5. Categorical-study stance, source ratings and communication. Panel (a) shows final categories by preset group. Panels (b) to (e) show 10-point trust, and panels (f) to (i) show group-interaction matrices. All panels use the categorical study and its 10-point trust scale. Axis labels EA, EGS and NR refer to the preset groups. Stance distributions, source-rating boxplots and group communication matrices for the categorical study.

Figure 6. Reported trust by condition and preset group under the categorical protocol. Each row contains ten agents. Filled boxes show quartiles, lines mark medians, diamonds mark means and dots show recorded scores. Colours identify the three preset groups. Twelve horizontal trust boxplots from the original 120 agent records, on the original ten-point scale.

Figure 7. Affect and central-square occupancy across 21 steps. Panels (a) to (d) report mean summed VAD category scores by preset group. Panel (e) counts residents present in the central square. Markers denote saved values and connecting curves are spline interpolants. Four affect-summary trajectories and the combined central-square occupancy panel.

Agent	P1 strength	P1 rank	P2 strength	P2 rank	P3 strength	P3 rank
Ava Ramires	53	2	50	2.5	28	8.5
Caleb Morris	30	9.5	47	5	43	3
Eleanor Finch	54	1	52	1	41	4
Emily Novak	30	9.5	49	4	32	6
Jade Herrera	39	8	41	9	35	5
Jonas Müller	40	7	44	7.5	60	2

Appendix

Table 21. Interaction strengths and ranks for every primary-network agent.

Node	P1 strength	P1 rank	P2 strength	P2 rank	P3 strength	P3 rank
Researcher	37	9.5	47	7	71	2
Ava Ramires	53	3	63	1	28	11
Caleb Morris	37	9.5	48	6	45	5
Eleanor Finch	63	1	62	2	63	3.5
Emily Novak	30	11	53	3	37	7.5
Jade Herrera	39	8	45	8.5	36	9

Appendix

Table 22. Researcher-inclusive strengths and average-tie ranks.

Explore similar work

Aug 25, 2026cs.CL

Belief Cascades Drive Persuasion in LLM Agent Networks

Multi-agent LLM systems increasingly debate answers, coordinate research, simulate users, and mediate information flows, making agent-to-agent persuasion a basic but undermeasured capability. We introduce a controlled testbed for studying how goal-directed persuaders shift elicited stances in networks of LLM agents grounded in real-world ego-network topologies. Across four LLM backbones, five graphs, and 55 policy statements, we find that persuasion dynamics depend on the interaction between topology, competition, topic, and model prior. Additionally, we show that direct exposure reliably predicts next-round stance change in competing runs, and peer relays carry smaller but measurable influence, showing that agents not assigned to persuade can still transmit persuasive force. Finally, analyzing post text alone misses important movement: planned strategies are only partly realized in executed messages, action choices can diverge from message content, and persuadees rarely state the stance shifts detected by probes. These results argue for evaluating multi-agent persuasion as a trajectory- and exposure-level process, using belief probes, exposure provenance, and action logs to identify who influenced whom and whether visible language reflects underlying stance movement.

Haoyi Qiu, Genglin Liu, Pranav Narayanan Venkit +4

University of California, Los Angeles · Salesforce AI Research

Persuasion in Language Models Multi-Agent LLM Systems LLM Agent Evaluation Belief Updating in LLMs Large Language Model-Based Social Simulation

Ask PDF

Sep 21, 2026cs.LG

Mind or Message? Auditing Theory of Mind in Multi-Agent Social Simulation

Language model agents are increasingly used to simulate social interaction, and the resulting transcripts read as though the agents understand one another. We ask whether that appearance rests on a model of the partner's mind or on the surface record of what the partner said. We build a social simulation in which both questions have exact answers: 40 multi-issue negotiations whose hidden preference weights and whose full Pareto frontier are known by construction. Two model families negotiate across 160 dyads, every transcript is frozen before any measurement, and 2880 counterfactual probes then hold the evidence byte identical while moving one factor at a time: the reader's own stake, the partner's tone, an identity label, and the order of recursion. The agents are socially fluent and economically poor. They reach agreement in 96.2% of dyads with 0 protocol failures, yet only 0.7% of deals land on the Pareto frontier, they leave 20.5% of the available joint value unclaimed, and they miss the one issue on which their interests are perfectly aligned in 76.6% of deals; on the frontier and on that aligned issue, a package drawn at random from the set both sides would accept does as well. The probes locate the failure. Swapping only the reader's own payoff sheet, while the partner's words and offers stay identical, moves the inferred top priority by 15.0 percentage points, which is egocentric projection rather than inference, while a tone rewrite moves it by 5.3 percentage points and an identity label by 0.0. Most tellingly, an agent predicts what its partner believes about it 72.5% of the time while that partner's belief is itself correct only 51.2% of the time: the agents track the conversation far better than they track the mind behind it.

Cong Li, Cheng Chen, Thomas Fung +2

DailyAdvance

Theory of Mind LLM Agent Evaluation Large Language Model-Based Social Simulation

Ask PDF

Sep 7, 2026cs.AI

From Simulated Citizens to Simulated Deliberation: Challenges in Representation and Interaction

Multi-agent LLM deliberation has been explored as a scalable way to simulate public deliberation. For such simulations to be informative, persona agents should reflect population opinion patterns and interaction should shape their conclusions. We evaluate whether LLM-based deliberation can meet these two conditions using census-grounded Korean personas debating real policy questions benchmarked against national surveys. Persona agents do not reliably reproduce population opinion patterns: responses are often far more concentrated and frequently reverse demographic differences in the human data. Deliberations nonetheless produce reasoned, reciprocal, and varied arguments alongside substantial stance movement. Yet much of this movement does not require peer exchange: sealed-monologue agents change position at similar rates and reach nearly the same final balance as full debates, while groups initialized with very different positions often converge to similar endpoints. Anchoring population-informed starting positions, meanwhile, sharply suppresses updating. Thus, population representation, argument generation, and interaction-driven opinion change do not necessarily go together. The simulations readily surface arguments on both sides, though whether they capture the diversity of human perspectives remains untested, leaving open a promising role for argument surfacing even as population simulation requires further validation.

Chaemin Jang, Junsik Min, Jaewoo Choi +9

1Korea Advanced Institute of Science and Technology (KAIST) · 2Seoul National University

Opinion Dynamics Human Behavior Simulation Multi-Agent Debate Large Language Model-Based Social Simulation Demographic Bias in Language Models

Ask PDF