AI Networks That Learned to Keep the Beat Mirror Human Rhythm Synchronization

Human beings are remarkably good at locking onto a rhythm. Tap your finger along with a song and your brain is performing a sophisticated computational feat: extracting a steady pulse from a complex acoustic signal, predicting when the next beat will arrive, and adjusting your movements on the fly when the tempo shifts. A new study published in Communications Psychology shows that artificial neural networks, trained with nothing more than reinforcement learning, can reproduce these human beat-synchronization dynamics with striking fidelity—offering a fresh window into how biological brains may accomplish the same task.

The research addresses one of the most enduring puzzles in music cognition. Humans do not merely react to beats after they occur; we anticipate them, tapping slightly ahead of or behind the pulse in ways that follow lawful, measurable patterns. Decades of psychophysical experiments have catalogued these patterns in meticulous detail, including the negative mean asynchrony—the tendency for people to tap a few tens of milliseconds before the actual beat—and the way synchronization error grows predictably as tempo increases. What has remained contested is what kind of internal mechanism could give rise to these behaviors.

Traditional computational models have approached the problem in two main ways. Event-based models treat beat tracking as a discrete prediction problem, estimating the time of the next beat from a sequence of previous onsets. Dynamical systems models, by contrast, embed rhythm perception in continuous oscillatory processes, with neural oscillators that entrain to external periodic input. Each family of models captures part of the behavioral data, but each requires engineers to hand-design key structural assumptions about how timing information is represented. The new work asks a different question: if a general-purpose learning system is simply rewarded for staying in sync, will it discover these human-like dynamics on its own?

To find out, the researchers built recurrent neural networks—flexible computational systems whose internal activity depends on both their current input and their own prior states—and trained them using reinforcement learning. Rather than being given labeled examples of correct taps or supervised gradient signals tied to human behavior, the networks received a simple scalar reward for minimizing the error between their produced taps and the underlying beat of rhythmic sequences. This setup mirrors the ecological situation faced by a human learner, who must discover synchronization strategies through feedback rather than explicit instruction about where the beat lies.

After training on a diverse repertoire of rhythmic patterns and tempos, the networks were tested under the same conditions used in classic human tapping experiments. The results were notable on several fronts. The trained networks exhibited a negative mean asynchrony comparable in magnitude to that observed in human participants, and they reproduced the characteristic scaling of synchronization variability with tempo. When the rhythmic stimulus changed speed or switched pattern mid-sequence, the networks showed phase-correction and period-correction behaviors that closely matched human adaptation curves, including the asymmetric way people correct small errors more readily than large ones.

Perhaps most intriguingly, the internal dynamics of the trained networks revealed interpretable structure. Analysis of their recurrent activity showed convergence toward low-dimensional trajectories that effectively encoded beat phase and period, resembling the oscillator-like states posited by dynamical theories of beat perception. The networks had not been told to build internal clocks, yet the pressure of the reward function drove them to develop timing representations functionally equivalent to those long hypothesized in the neuroscience of rhythm. This suggests that human-like synchronization behavior may emerge from general learning principles rather than from rhythm-specific neural hardware alone.

The findings carry implications that extend beyond music cognition. Beat synchronization is increasingly used as a clinical and developmental marker: individual differences in tapping performance have been linked to language development, reading ability, and a range of neurological conditions, including Parkinson’s disease and developmental dyslexia. If a simple reinforcement-trained network can capture the canonical signatures of human performance, researchers gain a tractable model system for probing which aspects of synchronization depend on sensorimotor learning, on neural oscillation, or on the statistics of the rhythms themselves. Perturbing the artificial networks—lesioning units, altering reward structure, or changing input statistics—offers a level of experimental control impossible with human participants.

The work also speaks to a broader debate in artificial intelligence and neuroscience about the explanatory power of trained neural networks as scientific models. Critics argue that deep learning systems can fit behavior without illuminating mechanism. This study pushes back on that concern by showing that a behaviorally constrained network not only matched human outputs but also reproduced interpretable internal dynamics aligned with theoretical predictions. In this sense, the network functions as a computational hypothesis: it demonstrates that the reward structure of a synchronization task is sufficient, in principle, to generate the observed suite of human tapping phenomena, narrowing the space of explanations that cognitive science needs to consider.

Limitations remain, as the authors acknowledge. Real-world beat perception unfolds in acoustically rich environments, with polyphonic music, expressive timing, and social coordination among multiple players, whereas the trained networks operated on more controlled rhythmic stimuli. Human synchronization also engages motor physiology—the biomechanics of the limb, spinal circuitry, and cortical motor areas—that a simulated tapping output does not capture. Extending the approach to multimodal, embodied settings will be an important next step, as will testing whether networks trained on different reward schedules produce different asynchrony profiles, which could help explain the substantial individual variability seen in human tapping data.

Even with those caveats, the study marks a compelling convergence between machine learning and cognitive science. It suggests that the human sense of the beat—so effortless that we rarely notice it—may be the natural solution to a general predictive-control problem, one that an artificial learner rediscovers when faced with the same objective. As reinforcement-trained models continue to reproduce facets of perception, action, and now rhythmic coordination, the boundary between engineered and biological intelligence grows both more interesting and more informative to study.

Subject of Research: Reinforcement-trained recurrent neural networks as computational models of human beat synchronization and sensorimotor timing

Article Title: Reinforcement-trained recurrent networks reproduce human beat-synchronization dynamics

Article References: Ommi, Y., Yousefabadi, M., & Cannon, J. (2026). Reinforcement-trained recurrent networks reproduce human beat-synchronization dynamics. Communications Psychology. https://doi.org/10.1038/s44271-026-00529-1

Image Credits: AI Generated

DOI: 10.1038/s44271-026-00529-1

Keywords: beat synchronization, recurrent neural networks, reinforcement learning, music cognition, sensorimotor timing, rhythm perception, neural oscillators, tapping behavior, computational neuroscience, predictive timing, Reinforcement-trained, recurrent