How do you compare the "thoughts" of a biological brain to the internal states of a silicon-based neural network? A human brain consists of billions of neurons firing in complex temporal patterns; a Large Language Model (LLM) consists of billions of floating-point weights organized in transformer blocks. They share no common physical substrate, no common "unit" of measurement, and no direct point-to-point mapping.
To solve this problem of incommensurability, researchers have converged on a framework called Representational Similarity Analysis (RSA). Instead of trying to align individual neurons with individual artificial units, RSA compares the relational geometry of how different systems organize information.
The Geometry of Information
The core intuition behind RSA is that while two systems might represent a "dog" using completely different physical signals, they might still agree on how a dog relates to a cat, a car, or a carrot.
As Marieke Mur, Nikolaus Kriegeskorte, and colleagues argued in Human Object-Similarity Judgments Reflect and Transcend the Primate-IT Object Representation, we can map these relationships into a shared representational space. If you measure the distance between the neural patterns for "dog" and "wolf," you will find they are closer together than the patterns for "dog" and "skyscraper."
By calculating the distances between all possible pairs of stimuli in a set, we create a Representational Dissimilarity Matrix (RDM). This matrix acts as a signature of the system’s internal logic. As Hamed Nili and Cai Wingfield detailed in A Toolbox for Representational Similarity Analysis, this abstraction allows us to compare disparate systems—monkeys, humans, and AI—by simply correlating their RDMs. This is known as "second-order isomorphism": the systems are similar because the relationships between their representations are the same.
Convergence in Vision and Action
One of the most striking successes of this approach is in the study of the visual system. Researchers like Daniel Yamins and James J. DiCarlo demonstrated in A performance-optimized model of neural responses across the ventral visual stream that when artificial networks are optimized for high-level object recognition, their internal representational geometry naturally begins to mirror the human ventral visual stream.
This alignment isn't just about static images. In Spatio-temporal convolutional neural networks explain human neural representations of action recognition, Andrea Tacchetti and Leyla Işık showed that models optimized for "invariance"—recognizing an action regardless of the actor or viewpoint—produce representations that most closely match human MEG data. The brain and the AI are arriving at the same geometric solutions to the same engineering problems.
Scaling Toward Human-Like Language
The same principles are now being applied to Large Language Models. In Large Language Models Converge on Brain-Like Word Representations, Jiaang Li and Anders Søgaard found that as LLMs scale up, their internal word geometries become increasingly similar to human fMRI patterns.
This suggests a "semantic convergence." Even when models are trained on different data or with different architectures, they seem to settle on a universal representational language. Daniel Son and colleagues explored this in Semantic Convergence: Investigating Shared Representations Across Scaled LLMs, using sparse autoencoders to show that models of different sizes independently discover the same conceptual building blocks.
Furthermore, Greta Tuckute and Evelina Fedorenko have shown in Language models align with brain regions that represent concepts across modalities that LLMs align best with "semantically rich" brain regions—those that represent the same concept whether it is seen in an image or read in text.
Attention as a Geometric Sculptor
Representational geometry is not static; it is dynamically reshaped by our goals. Samuel A. Nastase and Andrew C. Connolly demonstrated this in Attention selectively reshapes the geometry of distributed semantic representation.
When you look at a video of a dog, your brain creates a specific RDM. But if you are asked to focus on the dog's behavior rather than its species, your brain "stretches" the distances between different behaviors in its representational map, making them easier to distinguish. RSA allows us to quantify this "sculpting" of the internal map, showing how attention acts as a top-down filter on our geometric representations.
The Fragility of Alignment
While RSA is a powerful bridge, it is not without caveats. Recent work by Ansh Soni and Meenakshi Khosla in Conclusions about Neural Network to Brain Alignment are Profoundly Impacted by the Similarity Measure warns that our conclusions about "brain-likeness" are highly sensitive to the specific mathematical metrics used.
Moreover, Yantong Wang and Dawei Liang point out an "embodiment gap" in Cognitive Alignment Between Humans and LLMs Across Multimodal Domains. While AI might mirror our geometry for visual or linguistic concepts, it often fails to align with human representations of physical sensations like smell, taste, or causal reasoning—areas where the lack of a physical body makes the "geometry of thought" fundamentally different.