🧠 neuroscience

The brain prioritizes information that strengthens the structure of knowledge

This study reveals that the human brain prioritizes semantic information that strengthens the small-world structure of knowledge systems, a process mediated by the language network and ventromedial prefrontal cortex to enhance learning efficiency.

Original authors: Tikochinski, R., Pinhorn, I., Nahari, T., Duettmann, A., Sharot, T.

Published 2026-07-13

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Tikochinski, R., Pinhorn, I., Nahari, T., Duettmann, A., Sharot, T.

Original paper licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). ⚕️ This is an AI-generated explanation of a preprint that has not been peer-reviewed. It is not medical advice. Do not make health decisions based on this content. Read full disclaimer

Imagine your brain isn't just a hard drive storing facts like a dusty library. Instead, think of it as a bustling, super-connected city. In this city, every piece of knowledge is a building, and the roads connecting them are your thoughts. Some buildings are right next to each other (like "bread" and "butter"), while others are far apart (like "bread" and "space travel").

For a long time, scientists thought we just grabbed information randomly or only when we were curious about something specific. But a new study by researchers at University College London and MIT suggests something much cooler is happening: Your brain is a master city planner.

It seems your brain has a secret rule for deciding what information is worth paying attention to. It loves information that makes your mental city more efficient. Specifically, it craves facts that create "shortcuts" between distant neighborhoods while keeping local neighborhoods cozy and connected. In the language of network science, this is called a "small-world" structure.

The "Aha!" Factor

The researchers tested this by showing people 100 random facts, like "Bread was once used as currency." As the participants read each sentence, the scientists used a computer model to map out how that fact would change the "city" of human knowledge. They calculated a score called the Small-World Index (SWI).

Here is the magic: The people didn't just like these facts; they wanted them.

The Finding: The higher the fact's score for improving the city's layout (making it more connected and efficient), the more the participants said they liked it and the more they wanted to read similar facts.
The Proof: This wasn't just a feeling. When the researchers scanned the participants' brains with an fMRI machine, they saw the brain lighting up. The language network (the part that processes words) worked harder for these "efficient" facts. Even more exciting, the ventromedial prefrontal cortex (vmPFC)—the brain's "value center" that usually lights up for food, money, or art—also lit up. This suggests the brain treats a well-organized piece of knowledge like a delicious reward.

What It's NOT About

You might think, "Maybe people just liked the facts because they were surprising?" or "Maybe they liked facts that connected two things that were already very similar?"

The study explicitly says no to those ideas.

The researchers controlled for how surprising the facts were. Even if a fact wasn't shocking, if it improved the network structure, people still liked it.
They also checked if people just liked connecting similar things. Nope. The brain specifically valued the balance between keeping local groups tight and creating bridges to far-away ideas. It's not just about similarity; it's about the whole map getting better.

The "Virtual" Confirmation

To double-check that this isn't just a human quirk, the researchers asked a Large Language Model (an AI) to rate the same facts. While the AI doesn't have feelings or preferences of its own, it showed the exact same pattern. This suggests that the pattern is so deeply embedded in human writing and thinking that even an AI trained on our data picks up on it. It hints that this "efficiency preference" might be a fundamental rule of how we learn.

How Sure Are We?

The researchers are pretty confident, but they are careful with their words.

The Evidence: They ran three separate studies. One was a big online test with 93 people. The second was a brain scan study with 33 people. The third was a replication using a different set of 31 people and different facts. The results were the same every time.
The Simulation: The "city map" they used was a computer simulation based on Wikipedia entries. They admit this isn't a perfect, personalized map of every single person's mind (which would be huge and unique). However, even with this "proxy" map, the brain and behavior effects were strong and consistent.
The Conclusion: The paper suggests that the brain prioritizes information that strengthens the structure of what we know. It doesn't claim to have solved the mystery of learning entirely, but it offers a powerful new theory: We don't just want to know more; we want to know what makes what we already know work better.

So, the next time you find yourself diving deep into a topic that connects two totally different ideas, remember: your brain isn't just being curious. It's doing some serious urban planning, building bridges to make your mental world a faster, smarter, and more creative place.

Technical Summary: The Brain Prioritizes Information that Strengthens the Structure of Knowledge

Problem Statement
Human cognition achieves remarkable comprehension and innovation despite severe constraints on time, resources, and computational capacity. While humans rely on internal models of the world to predict outcomes, they cannot process all available information indiscriminately. The central problem addressed is identifying the strategic principles by which the brain prioritizes new semantic inputs. Specifically, the authors investigate whether the brain preferentially selects information that optimizes the structural organization of existing knowledge, hypothesizing that information reinforcing a "small-world" network topology is intrinsically valued.

Methodology
The study employed a multi-experiment approach combining behavioral psychology, large language model (LLM) analysis, and functional magnetic resonance imaging (fMRI).

Knowledge Network Construction: The researchers constructed a proxy knowledge network consisting of 500 concepts (169 from the experimental stimuli and 331 sampled from Wikipedia based on TF-IDF salience). Concepts were embedded into a high-dimensional semantic space using OpenAI's text-embedding-ada-002 model (representing full Wikipedia entries rather than single words). A weighted, undirected graph was generated where edges existed between concepts if their cosine similarity exceeded the 80th percentile.
Small-World Index (SWI) Calculation: The SWI was computed as the ratio of the network's clustering coefficient ( $C$ ) to its average shortest path length ( $L$ ), normalized against random graphs of equivalent size and density ( $SWI = (C/C_{random}) / (L/L_{random})$ ).
Stimulus Manipulation: For each of 100 factual sentences (e.g., "Bread was once used as currency"), the researchers simulated the integration of new information by strengthening the edge weight between the two key concepts in the sentence. The change in the Small-World Index ( $\Delta SWI$ ) was calculated as the difference between the network's state before and after this strengthening.
Experiment 1 (Behavioral): 93 participants rated 100 facts on "liking," "wanting" (interest in related facts), and "surprise." Linear mixed-effects models tested the association between these ratings and $\Delta SWI$ , controlling for concept similarity and demographic variables.
LLM Validation: OpenAI's GPT-4o was prompted to rate the same 100 facts on liking and wanting across 50 independent runs to determine if human-generated text reflects this preference pattern.
Experiment 2 (fMRI): 33 participants underwent fMRI while viewing the 100 facts. Parametric analyses modeled BOLD responses as a function of $\Delta SWI$ . Regions of Interest (ROIs) were defined via Neurosynth meta-analyses for "language" (bilateral STG, left IFG) and "reward" (vmPFC). Separate General Linear Models (GLMs) isolated the effects of changes in clustering ( $\Delta C$ ) and path length ( $\Delta L$ ).
Experiment 3 (Replication): An independent dataset (31 participants) involving 45 new two-concept sentences was analyzed. This replication utilized participant-specific functional language localizers rather than meta-analytic masks to confirm neural findings.

Key Results

Behavioral Preference: Participants significantly preferred facts that increased the small-world configuration of the knowledge network. Higher $\Delta SWI$ predicted greater "liking" ( $\beta = 0.052, p = .0002$ ) and "wanting" ( $\beta = 0.031, p = .0025$ ). This effect persisted after controlling for surprise ratings and the cosine similarity between concepts. Model comparisons indicated that the composite $\Delta SWI$ metric explained preference better than $\Delta C$ or $\Delta L$ alone.
LLM Alignment: GPT-4o exhibited the same preference pattern, with $\Delta SWI$ significantly predicting both liking and wanting ratings, suggesting this preference is embedded in human-generated text.
Neural Prioritization (Language System): The language network (bilateral STG and left IFG) showed increased BOLD activity proportional to $\Delta SWI$ . This response was driven by both increased clustering and reduced path length, with the composite $\Delta SWI$ showing a stronger effect than either component in isolation.
Neural Prioritization (Valuation System): The ventromedial prefrontal cortex (vmPFC), a core node of the brain's valuation system, showed increased activation for facts with higher $\Delta SWI$ . However, unlike the language network, the vmPFC response was primarily driven by the reduction in average path length ( $\Delta L$ ), not clustering ( $\Delta C$ ).
Whole-Brain Findings: Exploratory analyses revealed additional activation in the medial multi-demand network (dmPFC), left-lateralized language regions, and bilateral cerebellum/ventral temporal areas associated with semantic integration.
Replication: Experiment 3 successfully replicated the behavioral preference for $\Delta SWI$ and the neural prioritization in both the language network and vmPFC using a different stimulus set and individually defined ROIs.

Significance and Claims
The authors claim to identify a fundamental principle of information prioritization: the brain assigns value and resources to knowledge based on its capacity to improve the structural organization of what is already known. Specifically, the brain favors information that strengthens the small-world topology of semantic networks—a structure associated with efficiency, creativity, and comprehension.

The paper posits that this mechanism may explain how humans learn effectively from limited data. By preferentially engaging with information that optimizes the trade-off between local coherence (clustering) and global integration (short path lengths), the brain can build powerful mental models without processing all available data. The findings suggest that the "neural common currency" for value extends to the structural utility of information, not just primary or secondary rewards. Furthermore, the observation that LLMs reflect this preference implies that this optimization principle is deeply embedded in human knowledge representation. The authors conclude that the brain does not merely seek to know more, but seeks to know what makes existing knowledge more useful and structured.

The "Aha!" Factor

What It's NOT About

The "Virtual" Confirmation

How Sure Are We?

More like this