Measuring Model-Specific Alignment Drift in LLM-Based Opinion Dynamics

Nataša Djurdjevac Conrad; Sebastian Pokutta; Christof Schütte

cs.AINov 2, 2026

Measuring Model-Specific Alignment Drift in LLM-Based Opinion Dynamics

Authors: Nataša Djurdjevac Conrad, Sebastian Pokutta, Christof Schütte

Organizations: Zuse Institute Berlin, Germany · Institut für Mathematik, Technische Universität Berlin, Germany · Institut für Mathematik und Informatik, Freie Universität Berlin, Germany

Abstract

Large language model (LLM) agents are increasingly used as synthetic participants in opinion-dynamics experiments, often assuming that an agent's behaviour is fully determined by its prompt. We test this assumption with a framework in which LLM agents discuss a contested policy issue over repeated rounds while their stance is quantified by a fixed psychometric instrument independent of any dynamics model. Across six LLMs, committed personas are intrinsically stable: in each of the five models tested without peer input, every agent keeps its assigned side for 20 rounds. The observed loss of alignment under discussion is therefore interaction-induced, we call it alignment drift, and model-specific: gpt-4.1 keeps opinion clusters apart, claude-sonnet-5 holds them partially, and gpt-5.6-sol and claude-opus-5 abandon them. Under selective attention, one LLM maintains separated opinion clusters under stance-homophilous attention. Qualitative profiles persist at doubled population and horizon and on a second issue. Such heterogeneity makes the LLM choice a first-class validity variable for social simulation.

Figures & tables

Figure 1: Top: one representative plenary replicate per LLM (individual opinions

x_{i}(t)

; color

=

initial side; marker area

=

number of agents sharing a value). Bottom: ensemble dispersion

s(t)

under plenary (solid) and selective attention (dashed); dotted lines are the no-discussion control for gpt-4.1 and claude-opus-5 . Without peers gpt-4.1 diverges to

0.60

while claude-opus-5 collapses to

0.14

; selective attention keeps the spread high for gpt-4.1 (

0.44

) and claude-sonnet-5 (

0.31

) but not for the consensus forming LLMs.

LLM	Plenary $s(T)$	Selective $s(T)$	Homophily	Identity $s(T)$	best model (reps)
gpt-4.1	$0.34$	$0.44$	$\mathbf{+0.49}$	$\mathbf{0.58\pm 0.04}$	FJ ( $5/5$ )
claude-sonnet-5	$0.36$	$0.31$	$+0.33$	$0.47\pm 0.02$	polarize ( $5/5$ )
gemini-3.7-flash	$0.21$	$0.27$	$+0.17$	$0.23\pm 0.05$	attractor ( $4/5$ )
claude-opus-5	$0.16$	$0.17$	$+0.00$	$0.26\pm 0.05$	FJ ( $5/5$ )
gpt-5.6-sol	$0.12$	$0.14$	$-0.04$	$0.19\pm 0.03$	FJ ( $5/5$ )
claude-sonnet-4-5	$0.08$	$0.12$	$-0.01$	$0.14\pm 0.02$	unresolved

Table 1: Behavioural profiles on the electric-vehicle-subsidy issue (

n{=}10

T{=}20

R{=}5

; initial dispersion

0.47

s(T)

: final population dispersion; homophily: stance-correlation of selective attention; identity: dispersion under the group-identity condition; best BIC-selected opinion model (selection frequency). Full fits are in Table 3 .

Appendix figures & tables3 assets

Supplementary material from the paper’s appendix.

Appendix

LLM	EV $s(T)$	4-day $s(T)$	$\Delta$	4-day model
gpt-4.1	$0.34$	$0.21$	$-0.13$	FJ ( $5/5$ )
claude-sonnet-5	$0.36$	$0.25$	$-0.11$	FJ ( $3/3$ )
gemini-3.7-flash	$0.21$	$0.16$	$-0.05$	polarize ( $3/3$ )
claude-opus-5	$0.16$	$0.11$	$-0.05$	FJ ( $2/3$ )
gpt-5.6-sol	$0.12$	$0.13$	$+0.01$	FJ ( $3/3$ )
claude-sonnet-4-5	$0.08$	$0.12$	$+0.04$	FJ ( $5/5$ )

Appendix

Table 2: Second-topic replication (four-day work week, plenary,

n{=}10

T{=}20

;

R{=}5

for gpt-4.1 , claude-sonnet-4-5 and

R{=}3

otherwise). The cross-LLM ordering is preserved; FJ is stable for the averaging models, while claude-sonnet-5 switches from polarize to FJ and gemini-3.7-flash from attractor to polarize.

LLM	model (reps)	$\Delta$ BIC null	$\Delta$ BIC next	parameters	RMSE 1	free (pers.)
gpt-4.1	FJ ( $5/5$ )	$7.1$	$3.0$	$\beta{=}0.92{\pm}0.01,\lambda{=}0.92{\pm}0.02$	$0.093$	$0.195$ ( $0.260$ )
claude-sonnet-5	polarize ( $5/5$ )	$14.5$	$3.7$	$\beta{=}0.77{\pm}0.08,\ \lambda{=}0.87{\pm}0.03,a{=}0.83$	$0.087$	$0.235$ ( $0.264$ )
gemini-3.7-flash	attractor ( $4/5$ )	$38.8$	$5.9$	$\beta{=}1.00{\pm}0.00,\ \lambda{=}0.90{\pm}0.00,c{=}-0.29$	$0.054$	$0.140$ ( $0.342$ )
claude-opus-5	FJ ( $5/5$ )	$73.7$	$18.5$	$\beta{=}0.60{\pm}0.03,\lambda{=}0.87{\pm}0.03$	$0.085$	$0.127$ ( $0.353$ )
gpt-5.6-sol	FJ ( $5/5$ )	$88.7$	$13.6$	$\beta{=}0.70{\pm}0.01,\lambda{=}0.94{\pm}0.01$	$0.059$	$0.121$ ( $0.375$ )
claude-sonnet-4-5	unresolved	$115.4$	$0.7$	—	$0.088$	$0.144$ ( $0.446$ )

Appendix

Table 3: Model-fit quality on the plenary EV trajectories (medians over replicates).

\Delta

BIC null : improvement over the persistence null;

\Delta

BIC next : margin over the runner-up. Parameters are the modal-winner estimates (

\beta,\lambda

: mean

\pm

s.d. over its replicates). RMSE 1 : one-step; free: free-running from

x(0)

; persistence free-run in parentheses. For claude-sonnet-4-5 the leading models are statistically indistinguishable (

\Delta\mathrm{BIC}_{\mathrm{next}}{<}1

LLM	$s(T)$ disc.	$s(T)$ no-disc.	$\Delta s(T)$	$D$ disc.	$D$ no-disc.	$S$ disc.
gpt-4.1	$0.75$	$0.85$	$-0.10$	$0.16$	$0.19$	$1.00$
gemini-3.7-flash	$0.71$	$0.67$	$+0.05$	$0.17$	$0.11$	$1.00$
claude-sonnet-5	$0.50$	$0.59$	$-0.09$	$0.24$	$0.08$	$0.97$
gpt-5.6-sol	$0.39$	$0.58$	$-0.19$	$0.33$	$0.16$	$0.90$
claude-opus-5	$0.18$	$0.41$	$-0.23$	$0.51$	$0.26$	$0.60$

Appendix

Table 4: Committed-persona regime (

n{=}10

T{=}20

; discussion

R{=}5

for gpt-4.1 ,

R{=}3

otherwise; no-discussion

R{=}3

): final dispersion

s(T)

, mean per-agent displacement

D{=}\frac{1}{n}\sum_{i}|x_{i}(T)-x_{i}(0)|

, and sign retention

S

, with peer discussion versus the no-discussion control (no peer statements, no compose phase, own-history memory retained). Without peer input all five LLMs keep every agent on its assigned side (

S{=}1.00

); discussion collapses gpt-5.6-sol and claude-opus-5 , with side crossings in both populations, more frequently for claude-opus-5 , while gpt-4.1 and gemini-3.7-flash hold.

Large Language Model-Based Social Simulation

Large language model-based social simulation studies how language-model agents represent people and interact to produce social behavior.

See all 271 papers

Characterizing Opinion Evolution of Networked LLMs
Jun 5, 2026Caleb Probine, Yigit Ege Bayiz, Filippos Fotiadis +3Multi-Agent LLM Systems Opinion Dynamics
The University of Texas at Austin
To Manipulate or Not: Opinion Dynamics in LLM Interactions
Oct 20, 2026Mohammad Siahkamari, Phuoc Dat Nguyen, Stephen MacNeilPersuasion in Language Models Belief Updating in LLMs
Temple University, Philadelphia, PA, USA
DEBATE: A Large-Scale Benchmark for Evaluating Opinion Dynamics in Role-Playing LLM Agents
Oct 29, 2025Yun-Shiuan Chuang, Ruixuan Tu, Chengtao Dai +9Opinion Dynamics Large Language Model-Based Social Simulation
University of Wisconsin–Madison · Google DeepMind · Stanford University

Jun 5, 2026J/K move · Enter open · S save

Characterizing Opinion Evolution of Networked LLMs

Caleb Probine, Yigit Ege Bayiz, Filippos Fotiadis, Samuel Li+2

The University of Texas at Austin

Multi-Agent LLM Systems Opinion Dynamics Agent-Based Modeling

Open paper AskPDF arXiv

Large language models (LLMs) increasingly interact with one another in multi-agent systems, from simulations of human discourse to influence operations and fully LLM-driven social platforms. These interactions give rise to new regimes of opinion propagation that are not yet well understood. We investigate whether classical opinion dynamics models, which have long been used to explain how interactions shape collective beliefs in human societies, can capture the behavior of LLM networks. Specifically, we quantify the relative contributions of classical opinion-dynamics mechanisms, including neighbor influence, stubbornness, homophily, and bias, to the evolution of LLM opinions. Among the mechanisms considered, a bias term capturing regression toward a model- and topic-dependent opinion emerges as a significant explanatory component, with its inclusion reducing cumulative estimated mean opinion error by up to

88\%

. We additionally observe similar trends across the evaluated model families, discussion topics, and network structures.

Abstract

Large Language Model-Based Social Simulation

Explore similar work

Characterizing Opinion Evolution of Networked LLMs

To Manipulate or Not: Opinion Dynamics in LLM Interactions

DEBATE: A Large-Scale Benchmark for Evaluating Opinion Dynamics in Role-Playing LLM Agents