The work sits at the intersection of quantum information, nonequilibrium physics, and reinforcement learning. Around it, there are already studies on controlling quantum systems with learning, on measurement-induced phase transitions, and on game-theoretic models of quantum circuits. But most neighboring works either use full information about the state or operate in mean field or with a small number of qubits. Here, the agent controls a large many-body system with only a fraction of the information about its state, and this changes the physics: an information threshold emerges, below which control is lost.
The authors model a one-dimensional stabilizer circuit of N qubits. At each step, with probability p the agent applies an optimally disentangling two-qubit gate chosen by the strategy, and with probability 1-p a random Clifford gate. The strategy receives as input a tableau (stabilizer table) in the clipped gauge, from which a fraction 1-q of rows is randomly removed, mimicking partial observability. The agent is trained with Proximal Policy Optimization (PPO) using a transformer architecture, with a reward proportional to minus the total entanglement. Simulations were run on GPUs using JAX, with system sizes N=16, 32, 64, 128.
The learned strategies dramatically outperform greedy and random ones. With full information (q=1), they completely eliminate the phase transition: for any arbitrarily small p>0, the steady state has an area-law entanglement, and the volume-law phase shrinks to the point p=0. The mechanism is the creation of a sequence of bottlenecks: the agent selects several equally spaced bonds and maintains nearly zero entanglement on them, which splits the chain into subsystems and limits the maximum achievable entanglement. The entanglement profile takes the form of pyramids, self-similar when the system size is doubled. With partial information, a critical threshold $q_c$(p) is found: above it, area-law; below it, volume-law. As p increases, the required information fraction decreases. For p≥$p_c^{gree}$dy, the area-law persists even as q→0, because the agent effectively catches Bell pairs using the single available stabilizer.
Unlike neighboring works where agents acted in mean field or on small systems, here control over individual degrees of freedom is demonstrated in a system of 128 qubits with partial information. This is the first test that reinforcement learning can stabilize nonequilibrium phases with area-law entanglement in the thermodynamic limit, and that full state tomography is not needed. The work also shows that simple deterministic rules devised by humans fail against stochastic dynamics; real-time active feedback is necessary.
A natural continuation is to study critical properties near the information threshold $q_c$(p) and the critical point of the greedy strategy $p_c^{gree}$dy: critical exponents, dynamic scaling behavior. An interesting question is whether this transition is universal or model-dependent. Another direction is to transfer to circuits with random unitary gates (Haar-random), where no transition was previously found in the thermodynamic limit; can learned strategies create bottlenecks there as well? Also promising is extension to two-dimensional systems and other nonequilibrium phases, such as topological ones.
The results directly impact the development of quantum error correction protocols and quantum memory, where suppressing entanglement with noise is critical. The method could be adapted for state preparation and stabilization of topological phases in real quantum devices with limited measurement access.
Immediate next steps are numerical determination of critical exponents for the information transition and analysis of strategy robustness to hardware noise. It is also important to test transferability of learning to circuits with continuous gate parameters, approaching real quantum processors.
The work connects to the fundamental problem of quantum many-body physics: how information and measurements affect entanglement dynamics and the emergence of nonequilibrium phases. An open question is whether there exists a universal class of transitions driven by available information, and what role active feedback plays in the thermodynamics of quantum systems, echoing the Maxwell's demon paradox.
This section is addressed to the author of the paper and their colleagues. Neighbouring works were found by meaning-based search across our archive, not by word overlap. We do not review: we show the field around the work.
What the work does
The authors investigate the control of quantum entanglement in one-dimensional stabilizer circuits using reinforcement learning agents that act on partial information about the state. They train strategies that, given a sufficient fraction of observed stabilizers, transform the volume-law growth of entanglement into an area-law, creating entanglement 'bottlenecks.' They find a critical information threshold below which control is lost.
Where its strength is
The work demonstrates that trained strategies can control entanglement in systems up to 128 qubits using only partial information, and they outperform simple greedy rules. This opens the way to studying the role of information in nonequilibrium quantum phases.
Where it could go next
Worth trying
Why this may matter
Understanding the minimal information required to control quantum chaos could matter for designing quantum devices where full state tomography is often impossible, and for developing machine learning methods in many-body physics.
Nearby works in our archive
The archive is sparse around this work: the closest review is only 0.55, against a usual 0.64. It looks like few have drilled nearby.
For a system of 128 qubits, the trained agent uses only about 13 stabilizers out of 128 (at q=0.1), yet still maintains area-law entanglement, catching Bell pairs with a probability of 76%.
The idea of a demon that, knowing only part of the system's information, keeps it from chaos, resembles Maxwell's demon from science fiction, but here the demon is a neural network trained by trial and error.