When a sequence model is trained to predict the next token in a stream of board game moves or maze coordinates, it receives no explicit information about physical space. It sees no grid, no pixels, and no geometric coordinates—only a flat, linear sequence of discrete symbols. Yet, internally, the model does something remarkable: it reconstructs a functional, continuous map of the physical world.
Why does this happen? The emergence of spatial representations is not an accidental quirk of deep learning. Instead, it is a mathematically optimal strategy for data compression. To predict the future states of a system governed by physical constraints, a neural network must internalize the underlying geometry of that system.
From Linear Sequences to Metric Spaces
Consider a model trained on Othello game sequences. In Emergent World Representations: Exploring a Sequence Model Trained on a Synthetic Task, Kenneth Li et al. demonstrated that a GPT-style architecture spontaneously builds a functional 2D representation of the board. By intervening on the model's internal activations, the authors proved that this latent map is actively used to determine legal moves.
This phenomenon generalizes beyond board games. When trained on maze-solving tasks, as shown by Michael I. Ivanitskiy et al. in Linearly Structured World Representations in Maze-Solving Transformers, Transformers organize discrete coordinate tokens within their embedding space such that their geometric distance matches their physical proximity, such as Manhattan distance. Similarly, Anders Søgaard et al. in Revisiting the Othello World Model Hypothesis confirmed that across various architectures, embeddings of adjacent physical locations naturally cluster closer together than those of distant ones.
The intuition here rests on statistical overlap. If location $A$ is physically adjacent to location $B$, the legal transitions or subsequent tokens from $A$ and $B$ will share massive statistical commonalities. To minimize prediction error, the network projects these tokens onto a low-dimensional manifold where geometric distance mirrors relational proximity.
Time, Action, and Sensorimotor Loops
Space can also emerge from the relationship between time and action. In the physical world, an agent cannot instantly teleport from one room to another; it must move through continuous intermediate states. This temporal smoothness provides a powerful supervisory signal.
By optimizing for "temporal slowness"—the principle that meaningful environmental features change slowly over time compared to raw pixel fluctuations—networks can recover physical coordinates. In Learning Subgoal Representations with Slow Dynamics, Siyuan Li et al. demonstrated that latent spaces optimized for slow dynamics naturally reconstruct the underlying physical topology of an environment from high-dimensional images.
When an agent can act, the mathematical relationship between its actions and sensory transitions forces the latent space to adopt the global topology of the environment. In Homomorphism AutoEncoder --- Learning Group Structured Representations from Observed Transitions, Hamza Keurti showed that an autoencoder can autonomously discover complex global structures, such as a toroidal geometry, simply by modeling how actions transform states. Furthermore, Alban Laflaquière et al. in Unsupervised Emergence of Egocentric Spatial Structure from Sensorimotor Prediction proved that an agent can discover Euclidean space and metric regularity purely by learning to predict its future sensory states from motor commands.
The Pressure of Downstream Planning
Explicit geometric maps are often treated as a prerequisite for navigation. However, research shows that spatial mapping is an emergent solution to the problem of navigation itself.
When a model is trained end-to-end to navigate or plan, its internal representations are shaped by the utility of the map for the task. In Cognitive Mapping and Planning for Visual Navigation, Saurabh Gupta et al. demonstrated that optimizing a joint mapping and planning architecture allows latent spatial representations to emerge without explicit geometric labels. Because the network must predict path reachability, it learns to exploit statistical regularities, such as inferring the layout of unobserved rooms from visual cues like door frames.
Remarkably, this occurs even in the complete absence of vision. In Emergence of Maps in the Memories of Blind Navigation Agents, Erik Wijmans et al. showed that agents sensing only their own egomotion spontaneously develop internal metric occupancy grids. The memory of a recurrent network becomes a spatial map because maintaining a persistent tracking of boundaries and open spaces is the most efficient way to avoid obstacles and reach a goal.
Universal Geometric Principles
Why do these emergent representations often mirror biological structures, like the grid cells found in mammalian brains? The answer lies in pure geometry rather than biological happenstance.
In On Conformal Isometry of Grid Cells: Learning Distance-Preserving Position Embedding, Dehong Xu et al. proved that if a neural network is optimized to preserve local physical distances—a property known as conformal isometry—it mathematically converges on a hexagonal flat torus topology. This explains why hexagonal grid patterns emerge across diverse artificial architectures: it is the most efficient way to embed a continuous 2D metric space into bounded neural activations.
Whether a model is predicting the next token in a text stream, navigating a 3D environment, or predicting sensorimotor transitions, space emerges as a fundamental latent variable. Neural networks do not need to be told that the world is spatial; they discover geometry because it is the most elegant, compressed explanation for the structured data they are forced to predict.
Key Foundational Work
Emergence of Global Structure from Local Associations (1993) — contributed 2 concept elements to this direction.