Neural Synchrony Between Socially Interacting Language Models

Zhining Zhang; Wentao Zhu; Chi Han; Yizhou Wang; Heng Ji

cs.CLFeb 19, 2026

Neural Synchrony Between Socially Interacting Language Models

Authors: Zhining Zhang, Wentao Zhu, Chi Han, Yizhou Wang, Heng Ji

Organizations: Peking University · Eastern Institute of Technology, Ningbo · University of Illinois Urbana-Champaign

Abstract

Neuroscience has uncovered a fundamental mechanism of our social nature: human brain activity becomes synchronized with others in many social contexts involving interaction. Traditionally, social minds have been regarded as an exclusive property of living beings. Although large language models (LLMs) are widely accepted as powerful approximations of human behavior, with multi-LLM system being extensively explored to enhance their capabilities, it remains controversial whether they can be meaningfully compared to human social minds. In this work, we explore neural synchrony between socially interacting LLMs as an empirical evidence for this debate. Specifically, we introduce neural synchrony during social simulations as a novel proxy for analyzing the sociality of LLMs at the representational level. Through carefully designed experiments, we demonstrate that it reliably reflects both social engagement and temporal alignment in their interactions. Our findings indicate that neural synchrony between LLMs is strongly correlated with their social performance, highlighting an important link between neural synchrony and the social behaviors of LLMs. Our work offers a new perspective to examine the "social minds" of LLMs, highlighting surprising parallels in the internal dynamics that underlie human and LLM social interaction.

Figures & tables

Figure 1: An illustration of our analysis framework. (1) Two LLM agents engage in a social interaction, generating responses conditioned on their backgrounds, goals, and shared history. (2) Hidden representations are extracted at each turn from both LLMs. (3) Neural synchrony is measured by learning affine mappings to predict one agent’s future representations from the other’s current ones.

Figure 2: Example synchrony heatmaps of $A$ predicting $B$ . The horizontal and vertical axes denote source layers $l_{A}$ and target layers $l_{B}$ , respectively, with each cell showing the test-set $R^{2}$ score for the corresponding layer pair. Gray cells indicate negative test-set $R^{2}$ values. For each $l_{A}$ , the best-predicting $l_{B}$ is highlighted with a black box. Model names are abbreviated by omitting the suffix “Instruct” due to space constraints. Note that Llama-3.2-3B-Instruct has 28 layers, whereas the other models have 32 layers.

Figure 3: Examining synchrony of Mistral-7B-Instruct-0.3 predicting Llama-3-8B-Instruct under different conditions. Left: Experimental group with genuine interaction. Middle: Control group 1 ( w/o social engagement ), where one agent passively consumes the dialogue without actually engaging in the interaction. Right: Control group 2 ( w/o temporal proximity ), where source representations are paired with lagged future representations.

Figure 4: Neural synchrony declines under control settings. (a) Control group without social engagement: bar plots show the mean

\mathit{SyncR^{2}}

across models within each family. (b) Control group without temporal proximity: line plots show

\mathit{SyncR^{2}}

for all model pairs within each family as a function of lag

k

(

k\!=\!0

is the experimental group). Model names are abbreviated (see Appendix F for the full name correspondence). Error bars denote the standard errors.

Figure 5: Neural synchrony under different relationship closeness.

\mathit{SyncR^{2}}

is aggregated over all model pairs.

Figure 6: Correlation between neural synchrony and social performance across different family types. The

x

-axis denotes the average social performance of the LLM pair in Sotopia across all scenarios, evaluated by gpt-oss-120b which aligns closely with human evaluation. The

y

-axis shows the measured synchrony score

\mathit{SyncR^{2}}

of the two models. Each point is annotated with abbreviated model names (the full name correspondence is provided in Appendix F ). The results show a clear linear relationship between neural synchrony and social performance, where a general trend is that more synchronized models achieve better social performance.

Model Family	IFEval	MuSR
Mistral	0.81	0.92
Cross	0.71	0.89
Llama	0.27	0.99

Table 1: Partial correlations between neural synchrony and social performance after controlling for different abilities. Rows correspond to model family types, and columns correspond to the datasets used for evaluation and control.

Appendix figures & tables9 assets

Supplementary material from the paper’s appendix.

Appendix

Figure 7: Pearson correlation between

\mathit{SyncR^{2}}

(computed under the top-

k

setting) and social performance as a function of

k

. The x -axis shows the number of top predictive matches ( k , up to 32 layers). The y -axis shows the resulting Pearson correlation between

\mathit{SyncR^{2}}

and social performance. Marker size indicates statistical significance, while shape and color denote the model family.

Figure 8: Comparison of

\mathit{SyncR^{2}}

computed with and without clamping negative values. Labels on the x -axis correspond to model pairs (model names are abbreviated; see Appendix F for the full name correspondence), grouped by their model family. The upper panel reports

\mathit{SyncR^{2}}

values; the y -axis is truncated at –30 to exclude extreme outliers for visibility. The lower panel reports the standard errors of

\mathit{SyncR^{2}}

on a logarithmic scale.

Figure 9:

\mathit{SyncR^{2}}

as a function of dataset sample size for training affine transformations. Error bars indicate standard errors.

Model Pair	Original	New
Mistral-7B-Instruct-v0.2 & Mistral-7B-Instruct-v0.2	0.46 ± 0.01	0.47 ± 0.02
Mistral-7B-Instruct-v0.2 & Llama-3-8B-Instruct	0.55 ± 0.02	0.58 ± 0.02
Llama-3-8B-Instruct & Llama-3-8B-Instruct	0.75 ± 0.01	0.74 ± 0.05

Appendix

Table 2: Comparison of

SyncR^{2}

under original and new sampling implementations for the train and test split. In the new sampling, we guarantee that no same persona pair appears in both sets.

Abbreviation	Full Name
M1	Mistral-7B-Instruct-v0.1
M2	Mistral-7B-Instruct-v0.2
M3	Mistral-7B-Instruct-v0.3
L2	Llama-2-7B-Instruct
L3	Llama-3-8B-Instruct
L3.2	Llama-3.2-3B-Instruct

Appendix

Table 3: Correspondence between abbreviations and full model names.

Figure 10: Distribution of the difference between the scores given by humans and gpt-oss-120b .

Dim.	GPT-4	gpt-oss-120b
Secret	0.22 ∗∗	0.46 ∗∗
Knowledge	0.33 ∗∗	0.28 ∗∗
Social Rules	0.33 ∗∗	0.22 ∗
Believability	0.45 ∗∗	0.50 ∗∗
Relationship	0.56 ∗∗	0.45 ∗∗
Financial Benefits	0.62 ∗∗	0.38 ∗∗

Appendix

Table 4: Pearson correlation coefficients and

p

-values between different LLM evaluators’ evaluation and human judgment on models’ output among different dimensions. Strong and significant correlations are in blue.

Model Pair	Affine	Nonlinear
Mistral-7B-Instruct-v0.2 & Mistral-7B-Instruct-v0.2	$0.46\pm 0.01$	$0.47\pm 0.01$
Mistral-7B-Instruct-v0.2 & Llama-3-8B-Instruct	$0.55\pm 0.02$	$0.60\pm 0.01$
Llama-3-8B-Instruct & Llama-3-8B-Instruct	$0.75\pm 0.01$	$0.59\pm 0.01$

Appendix

Table 5: Comparison of nonlinear and affine transformations in measuring neural synchrony.

Figure 11: Decoding action and emotion distributions from current, previous, and next agent representations. The y-axis indicates test-set KL divergence (lower is better). Baselines are computed using the average emotion/action distributions from the training set. Results are averaged over three random seeds, with error bars showing standard errors.

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 257 papers

Jun 29, 2026

Collective cooperation without individual fidelity in LLM agents

Large language models (LLMs) are increasingly used as agents in simulations of social systems, yet it remains unclear when their behavior can be interpreted as a faithful proxy for human decision-making. Here we test LLM agents against a direct empirical benchmark: a large-scale networked Prisoner's Dilemma experiment with human participants. Using the same interaction protocol, payoff structure, and network topologies, we compare nine open-weight LLMs with the human data. The selected model reproduces several macro-level features of cooperation dynamics, including the early decline and later stabilization of cooperation. This aggregate agreement, however, does not extend uniformly to finer levels of behavior. LLM populations underestimate individual-level heterogeneity and generate conditional cooperation patterns that differ from those observed in humans. Adding a fraction of random agents improves some aspects of micro-level agreement, but does not remove the mismatch in decision rules. These findings reveal a macro--micro dissociation in LLM-based social agents: collective outcomes can appear human-like even when the underlying behavioral distributions and mechanisms are not. They suggest that validating LLM agents as human surrogates requires comparisons across aggregate dynamics, individual heterogeneity, and context-dependent decision rules, rather than outcome-level agreement alone.

Henrique Ferraz de Arruda, Carlos Gracia Lázaro, Alberto Aleta +1

ARAID Foundation, Zaragoza, Spain. · Institute for Biocomputation and Physics of Complex Systems, University of Zaragoza, Zaragoza, Spain. · Universidad San Jorge (USJ), Zaragoza, Spain. +1

Multi-Agent LLM Systems Human Behavior Simulation LLM Agent Evaluation Large Language Model-Based Social Simulation

Ask PDF

Sep 22, 2025

The PIMMUR Principles: Ensuring Validity in Collective Behavior of LLM Societies

Large language models (LLMs) are increasingly used to simulate human collective behavior, yet claims that such simulations are human-like remain largely untested. We conducted a systematic audit (pre-registered on OSF) of LLM-based social simulations across four databases (Scopus, IEEE Xplore, ACM Digital Library, and arXiv). Across 576 studies reported in 350 recent papers, we applied six methodological evaluations: agent Profile, Interaction, Memory, Minimal-Control, Unawareness, and Realism (PIMMUR). Coding every study against pre-specified rules, we revealed that PIM were met more often than MUR. Frontier LLMs correctly identified the underlying social experiment in 65.2% of cases, and 50.6% of prompts imposed constraints that pre-determined the outcome. These compliance rates are upper bounds, because incomplete methodological reporting (for example, unreleased prompts) limits the available evidence. Reproducing five representative experiments (e.g., opinion dynamics), we found that reported collective phenomena often vanish or reverse once PIMMUR principles are enforced, indicating that many "emergent" behaviors are methodological artifacts rather than genuine social dynamics. Current LLM simulations may therefore capture model-specific biases rather than universal features of human social behavior, raising concerns about their use as scientific proxies for human society.

Jiaxu Zhou, Jen-tse Huang, Xuhui Zhou +5

CUHK · JHU · CMU +3

LLM Evaluation LLM Auditing Human Behavior Simulation Agent-Based Modeling Large Language Model-Based Social Simulation

Ask PDF

May 5, 2026

Can LLMs Emulate Human Belief Dynamics?

Can LLMs simulate how humans form and change beliefs in social networks? We put this to the test by replicating an established study on belief dynamics, evaluating 12 LLMs across multiple model families and parameter sizes. The answer is a clear no, and in systematic ways. LLMs fail to capture initial human belief distributions and tend to be overall more conformist than humans, shifting their responses to align with those around them. They also take a nuanced approach to emulating human homophilic tendencies within networks. Our findings carry a double payoff: they highlight fundamental properties of LLM behavior, and they raise a sharp warning against deploying LLMs as human proxies in social simulations.

Adiba Mahbub Proma, Neeley Pate, James N. Druckman +3

Department of Computer Science University of Rochester · Department of Political Science University of Rochester · Department of Physics and Astronomy University of Rochester

Opinion Dynamics Normative Conformity in Language Models Human Behavior Simulation Large Language Model-Based Social Simulation

Ask PDF

Abstract

Large Language Model-Based Social Simulation

Explore similar work

Collective cooperation without individual fidelity in LLM agents

The PIMMUR Principles: Ensuring Validity in Collective Behavior of LLM Societies

Can LLMs Emulate Human Belief Dynamics?