Stevens Researchers Reveal How AI Models Mimic Human Social Reasoning

Theory of Mind: the human advantage

Theory of Mind allows people to infer and reason about the mental states of others, enabling us to anticipate their behaviors. Humans typically develop this capacity around age four, and the brain performs such reasoning effortlessly and with remarkable efficiency. “For a human brain, it’s a very easy task,” says Zhaozhuo Xu, Assistant Professor of Computer Science at the Stevens Institute of Technology’s School of Engineering. “And while doing so, our brains involve only a small subset of neurons, so it’s very energy efficient,” adds Denghui Zhang, Assistant Professor in Information Systems and Analytics at Stevens’ School of Business.

Contrasting human cognition and large language models

Large language models (LLMs), the focus of Zhang and Xu’s research, differ fundamentally from human cognition. While inspired by principles from neuroscience and cognitive science, these AI systems rely on artificial neural networks that learn statistical patterns from massive text corpora, rather than forming symbolic or experiential understanding. This architecture gives LLMs immense computational power but limits their efficiency.

Regardless of a task’s complexity, LLMs activate nearly all their internal parameters to generate responses. “When we humans evaluate a new task, we activate a very small part of our brain, but LLMs must activate pretty much all of their network,” Zhang explains. “They compute a lot of things they don’t need, which makes the process very inefficient.”

Mapping social reasoning in large language models

To address these inefficiencies, Zhang and Xu initiated a multidisciplinary collaboration to explore how LLMs perform social reasoning and how they might emulate the energy-efficient mechanisms of the human brain. Their study, How Large Language Models Encode Theory of Mind: A Study on Sparse Parameter Patterns, was published in Nature Partner Journal on Artificial Intelligence on August 28, 2025.

The researchers discovered that LLMs rely on a small, specialized subset of internal connections to handle social reasoning tasks, mirroring, to some extent, the brain’s selective activation of neurons. They further found that a model’s ability to reason about others’ beliefs depends heavily on how it represents word positions through a mechanism known as rotary positional encoding (RoPE). This encoding determines how the model tracks relationships among words and ideas, guiding its internal “attention” during reasoning about social interactions.

Key findings: structured “beliefs” within LLMs

“In simple terms, our results suggest that LLMs use built-in patterns for tracking positions and relationships between words to form internal ‘beliefs’ and make social inferences,” Zhang explains. The researchers’ analysis reveals that these models exhibit a rudimentary capacity for perspective-taking based on how they process and integrate positional information, an insight that bridges computational modeling and cognitive neuroscience.

Toward more efficient and human-like AI

By identifying how LLMs form and manipulate internal “beliefs,” the researchers believe it may be possible to design AI systems that operate with greater energy efficiency. “We all know that AI is energy expensive,” says Xu. “If we want to make it scalable, we have to change how it operates. Our human brain is very energy efficient, so we hope this research encourages efforts to make LLMs activate only a subset of parameters relevant to a specific task. That’s an important argument we want to convey.”

Bridging neuroscience and artificial intelligence

The findings represent an important step toward developing sparser, more efficient AI architectures that selectively engage their computational resources, akin to the modular organization of the human brain. By illuminating the internal mechanisms through which LLMs approximate social reasoning, Zhang and Xu’s work not only advances our understanding of artificial cognition but also offers new directions for building sustainable, scalable AI systems.