Neural Synchrony Between Socially Interacting Language Models

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.

💡 Research Summary

The paper “Neural Synchrony Between Socially Interacting Language Models” investigates whether large language models (LLMs) exhibit an analogue of human inter‑brain synchrony (IBS) when they engage in multi‑turn social dialogues. Drawing inspiration from neuroscience findings that human brain activity aligns during cooperative or communicative interactions, the authors propose a novel metric, SyncR², to quantify the alignment of internal representations (hidden states) between two interacting LLM agents.

Methodologically, the authors use the SOTOPIA simulation environment, where two agents are assigned distinct personas, goals, and background information and must negotiate or cooperate to achieve their objectives. At each dialogue turn, after generating a response, the hidden state at the final token of the prompt is extracted for each model, yielding a sequence of layer‑wise vectors h_A^t and h_B^t. To assess synchrony, they train ridge‑regression affine mappings from the representation of model A at turn t (source) to the representation of model B at turn t + 1 (target). The coefficient of determination (R²) on a held‑out test set measures how well A’s current state predicts B’s next state. For each source layer, the best‑predicting target layer is selected, negative R² values are clamped to zero, and the resulting scores are averaged across all source layers to produce SyncR²(A→B). The bidirectional SyncR²(A,B) is the mean of SyncR²(A→B) and SyncR²(B→A).

Two crucial control conditions are introduced. Control 1 replaces one active agent with a “passive” reader that only consumes the dialogue without generating responses; SyncR² drops dramatically, indicating that genuine engagement is required. Control 2 introduces temporal lag k (k ≥ 1) by pairing A’s current representation with B’s representation k turns later; synchrony declines sharply with increasing lag, confirming that the metric captures short‑term, temporally proximal dynamics rather than static similarity from shared context.

Experiments involve six open‑source LLMs (e.g., Llama‑3‑8B, Mistral‑7B‑Instruct) across 21 model pairs of varying architecture and size. The authors compute SyncR² for each pair during a suite of social tasks (negotiation, persuasion, coordination) and correlate the scores with behavioral performance metrics such as task success rate and reward. They find a robust positive correlation (Pearson r ≈ 0.68, p < 0.001), which persists after controlling for confounding factors like instruction‑following ability and long‑context reasoning. Notably, model size alone does not predict synchrony; instead, the degree to which one model’s internal state anticipates its partner’s next state is the key driver of social success.

The contributions are threefold: (1) introduction of a principled framework for measuring “neural synchrony” between interacting LLMs, bridging concepts from cognitive neuroscience to AI; (2) rigorous validation that the metric reflects both social engagement and temporal proximity, distinguishing genuine interaction from mere co‑exposure; (3) empirical evidence that higher representational synchrony predicts better social task performance across diverse model families.

By demonstrating that LLMs develop an internal, predictive alignment during dialogue—mirroring the functional role of human IBS—the work opens a new avenue for studying “social minds” in artificial agents. It suggests that future multi‑agent systems could be optimized by explicitly fostering representation‑level synchrony, potentially improving coordination, cooperation, and emergent collective intelligence in AI‑driven social environments.