LLM Mind-Reading Discovery May Lead to Better Performance

New study uses cognitive science to pinpoint how LLMs acquire theory of mind.

Posted November 20, 2025 | Reviewed by Margaret Foley

The human brain vastly outperforms artificial intelligence (AI) when it comes to energy efficiency. Large language models (LLMs) require enormous amounts of energy, so understanding how they “think" is a key competitive advantage toward lower computing costs and better overall performance. New research identifies how LLMs develop social reasoning capabilities and form theory of mind (ToM), an important discovery that may lead to more energy-efficient AI in the future.

“By illuminating the structural underpinnings of social intelligence in AI, our study bridges the gap between deep learning, cognitive science, and AI ethics ,” wrote the study's corresponding authors, Denghui Zhang and Zhaozhuo Xu at Stevens Institute of Technology, along with co-authors Heng Ji at University of Illinois Urbana-Champaign, Zirui Liu at University of Minnesota Twin Cities, Wentao Guo at Princeton University, and Yuheng Wu at Stanford University.

The term “theory of mind” was coined in 1978 by David Premack and Guy Woodruff at the University of Pennsylvania with the publication of their paper “Does the chimpanzee have a theory of mind?” in the journal Behavioral and Brain Sciences . In psychology, theory of mind refers to the capability to understand that one’s emotions, perceptions, intentions, thoughts, beliefs, and desires are different from others', an ability that is essential for handling social and emotional situations, as well as communications. In short, ToM is the ability to perform mind-reading , also known as mentalism. Humans develop this skill around the ages of 4 to 5, according to the Encyclopedia on Early Childhood Development .

LLMs, a subset of AI machine learning, have the troubling characteristic of lacking transparency and largely being a “black box.” Artificial neural networks are opaque when it comes to understanding how they arrive at their predictions and conclusions. Any progress towards shining a light on the inner workings of deep neural networks can not only translate into monetary benefits and higher quality output but also improve transparency, which is key in the realm of AI ethics.

“By exploring how LLMs develop the ability to infer mental states, we can better align LLM systems with human social cognition , fostering more trustworthy and interpretable interactions,” the researchers wrote.

To conduct their study, the team of scientists focused on the underlying mechanisms of how LLMs form theory of mind, particularly on the AI model’s parameters. The strategy of the researchers was to focus on the models' parameters to see which are highly responsive and impact theory-of-mind performance.

To do this, they created a method to detect patterns in LLM with sparse parameter sensitivity, where most of the values are approaching or equal to zero using mathematical formulas.

“Perturbing as little as 0.001% of model parameters leads to significant changes in ToM capabilities,” the researchers discovered.

Moreover, the team not only found the patterns in sparse parameters that are extremely responsive to theory-of-mind abilities but also found their connection to positional encoding. In models that contain Rotary Position Embedding (RoPE), the researchers discovered that those parameters that are highly responsive to ToM are tightly connected to the positional encoding module. In LLMs, positional encodings, also known as positional vectors, enable the tracking of each word’s position in a sequence. This tracking mechanism captures context and meaning. The order of words in a sequence impacts meaning. To illustrate the importance of positional encodings, the sentence “Fred feeds fish” has a quite different meaning than “Fish feeds Fred.”

Interestingly, the scientists found that the parameters that are highly responsive to ToM affect the model’s attention mechanisms that, if triggered, change the attention mechanism, and can diminish the model’s language comprehension and theory-of-mind abilities.

The pioneering researchers applied cognitive science and theory of mind at the parameter level to help understand the inner workings of AI large language models and uncovered important insights. They found that specific parameters that influence theory of mind tasks also influence the model’s language comprehension and context. The findings suggest that theory of mind reasoning in LLMs is emergent and arises from interactions between components.

“As LLMs continue to evolve, understanding how they acquire, encode, and manipulate social reasoning will be essential for ensuring their transparency, reliability, and alignment with human values,” the researchers concluded.

Share this post Facebook Bluesky Linkedin Email

There was a problem adding your email address. Please try again.

By submitting your information you agree to the Psychology Today Terms & Conditions and Privacy Policy

Cami Rosso writes about science, technology, innovation, and leadership.

Get the help you need from a therapist near you–a FREE service from Psychology Today.

This article is part of the Bringwise Psychology Journal — daily insights on human behavior, mental health, and personal growth.