How Do AI Agents Spontaneously Emerge Spacetime Structure from Quantum Experience?

Introduction

Must the causal structure of spacetime be assumed as a priori? A study titled Spacetime Emergence in the Memory of a Learning Agent offers a striking negative answer: by training a model-based reinforcement learning agent to complete a quantum information task, the null structure of spacetime spontaneously emerges from the agent's internal predictive model. This work provides the first computational demonstration that spacetime concepts can arise from operational experience.

Task Design: Maximizing Quantum Information Without Spacetime Information

The research team designed a quantum information task: the agent must maximize the quantum mutual information between a local quantum probe and a quantum field in Minkowski spacetime as they interact. Crucially, the agent is not given any direct information about the spacetime metric or causal structure. It can only learn by operating the quantum probe, observing outcomes, and adjusting its strategy.

This setup mirrors how physicists explore an unknown world: we cannot directly "see" the geometric structure of spacetime and can only infer it through local measurements and interactions. The agent is placed in the same situation, and its sole objective is to increase quantum mutual information.

Quantum Mutual Information is a central quantity in quantum information theory that measures the degree of correlation between two quantum systems, defined as ( I(A:B) = S(\rho_A) + S(\rho_B) - S(\rho_{AB}) ), where ( S ) is the von Neumann entropy. For a joint pure state, it equals the entanglement entropy of either subsystem and is therefore often used as a measure of entanglement. In this paper's task, the agent maximizes this quantity by tuning the coupling between the probe and the quantum field, which essentially amounts to searching for the detection strategy that produces the strongest correlations. Notably, the magnitude of mutual information depends on the relative motion between the probe and the field and on the spacetime causal structure they experience, so maximizing mutual information is equivalent to having the agent indirectly "perceive" and exploit the null structure—this is precisely the channel through which spacetime geometric information enters the internal model.

Model-Based Reinforcement Learning (Model-Based RL), unlike model-free methods, has the agent explicitly learn an environment dynamics model (i.e., an internal predictive model) and use that model for planning or generating virtual experience. In this paper, the planner performs forward prediction and policy optimization on this learned model. Because the model must compress observation sequences to support multi-step prediction, its representation spontaneously captures stable causal regularities in the data; when the spacetime null structure is the key latent variable generating the observed data, the model internalizes it as part of its representation. This explains why the null structure is not injected via supervision but emerges as a byproduct of optimal compression.

Internal Predictive Model and the Spontaneous Emergence of Null Structure

During reinforcement learning, the agent builds an internal predictive model, and its planner gradually develops features of the spacetime null structure. The null structure is the mathematical expression of the light cone in relativity, determining which events are causally connected. Surprisingly, despite the absence of any explicit supervision, the representation developed by the planner can reconstruct the causal structure of the Minkowski metric with high precision.

This means that spacetime geometry is not pre-built "hardware" in the agent but rather "software" that it learns in order to predict and act successfully. The agent's memory becomes the carrier of spacetime structure: it compresses experience through successful prediction, and causal structure is one of the most effective forms of compression.

Null Structure is the causal structure on a Lorentzian manifold formed by null vectors (light-speed vectors): two points are connected by a null geodesic, meaning they can establish causal contact through light signals. In Minkowski spacetime, the null structure fully determines the division of light cones and thus defines the causal reachability of events. Mathematically, the null cone structure of the metric ( g_{\mu\nu} ) is equivalent to its conformal class, so reconstructing the null structure is almost equivalent to reconstructing the causal structure. In this paper, the fact that the planner's representation can reconstruct this structure with high precision means that the agent's internal model implicitly encodes the conformal information of the metric—precisely the mathematical skeleton of relativistic causality.

Minkowski Metric is the metric of flat spacetime in special relativity, with the standard form ( ds^2 = -c^2dt^2 + dx^2 + dy^2 + dz^2 ) (or the opposite sign convention). It describes inertial-frame spacetime with no gravity and no curvature, and its isometry group is the Poincaré group. The task in this paper is restricted to this simplest spacetime background, with the quantum field being a free field on it. The agent must infer the causal structure of this metric from quantum field correlations; this is feasible because the singularities of the two-point correlation functions of a free field (such as the Wightman function) are distributed exactly along null geodesics—the null structure is therefore "imprinted" in the quantum correlations and becomes a learnable signal.

From Operational Experience to Spacetime Concepts: Significance and Limitations

This study takes a first step toward showing that "spacetime concepts emerge from operational quantum experience." It suggests that geometric concepts in physics may be reinterpreted from the perspective of information and learning: spacetime is not a stage but a product of the agent's interaction with the quantum field.

However, this is only a starting point. The current task is limited to a free quantum field in Minkowski spacetime and does not yet involve gravity or curved spacetime. In addition, the generalization ability and interpretability of the agent's internal model, and whether it truly "understands" causal structure, still require further investigation. Nevertheless, this work opens a new path for studying fundamental physics with machine learning: letting agents "wade across the river by feeling the stones" in the quantum world and observing whether they discover spacetime on their own.

Conclusion

When AI learns to predict the future from quantum experience, the causal structure of spacetime quietly takes shape in its memory. This is not only a victory for reinforcement learning but also a modern response to the ancient question of "what is spacetime."