Prediction is a core brain function: we constantly anticipate what we'll see next. A new study from researchers at Radboud University (Linda Ariel Ventura, Victoria Bosch, Tim C. Kietzmann, and Sushrut Thorat) shows that a simple predictive network can learn to bind objects to their locations—a key part of how we build mental maps.
The Research
In a minimal 'in-silico' setting, the team trained a recurrent neural network to sample tokens from 2D scenes. The network's task was to predict the next token given its current input and a simulated eye movement (saccade). On scenes it had never seen, prediction accuracy improved over the sequence, indicating 'in-context learning.'
Decoding analyses revealed that the network developed path integration—tracking its own movements—and dynamically bound token identity to position, just like our brains do when we navigate. Even more intriguing, interventional analyses showed that it could learn new bindings late in a sequence, and even for out-of-distribution pairings. These findings were accepted for presentation at CogSci 2026.
Why It Matters
This work offers a mechanistic account of how structured world models emerge from prediction. For human cognition, it suggests that our ability to remember where things are is closely tied to predicting what comes next. This has practical implications: when you actively anticipate what you'll see or do, you're not just preparing—you're building and refining your cognitive map.
What You Can Do
To strengthen your own object-location memory, try actively predicting your route before you walk it, or imagine the next step in a task before you do it. Such predictive engagement may help your brain bind information more flexibly and efficiently.
Source: arXiv q-bio.NC
Curious about your own brain? Take our free adaptive IQ test or try 306 brain training levels.