The animal at the center of the study is the mormyrid electric fish, a creature that generates its own weak electric field and senses the world through distortions of that field. Every time the fish fires its electric organ, it also sends a copy of the motor command, known as the corollary discharge, to a hindbrain structure called the electrosensory lateral line lobe, or ELL. This structure, architecturally reminiscent of the cerebellum, uses the corollary discharge to generate a prediction of the sensory input that the fish’s own discharge will produce. That prediction is then subtracted from the actual sensory stream, a process scientists call negative imaging. The result is elegant: predictable self-generated signals vanish, while unexpected signals, such as the electrical signature of nearby prey or a communicating neighbor, stand out in sharp relief.

What makes the fish such a powerful model is that this cancellation is learned continually. When the fish’s environment changes, for example when the amplitude of its electric organ discharge shifts, the circuit adjusts within minutes, and it does so without erasing previously learned predictions. This is a continual learning problem, the biological counterpart of a challenge that plagues artificial neural networks, where new training catastrophically overwrites old memories. Previous work from the same laboratories had shown that learning in the ELL is distributed across multiple synaptic sites, but the detailed wiring needed to explain how those sites cooperate had never been mapped.

To build that map, the team reconstructed a volume of the ELL at a resolution of 4 by 4 by 30 nanometers using serial section electron microscopy, a dataset generated in Jeff Lichtman’s laboratory at Harvard and segmented with automated tools developed at Google. The researchers painstakingly proofread and classified hundreds of neurons, including more than one hundred MG cells and nearly six hundred SG cells, tracing their axons and dendrites and annotating thousands of synapses. Every reconstructed connection was catalogued, allowing the team to ask not just which cells talk to which, but whether the pattern of that conversation is structured or random, and what such structure implies for computation.

The first major discovery concerns how sensory information enters the circuit. Theoretical models of cerebellum-like learning require that the instructive signals driving synaptic plasticity be conveyed through specific, separable pathways. The connectome revealed exactly that: sensory input reaches the output neurons of the ELL through both inhibitory and disinhibitory routes, carried by distinct classes of interneurons whose connectivity is far from arbitrary. These pathways fulfill long-standing theoretical requirements for instructing anti-Hebbian synaptic plasticity, the temporally asymmetric learning rule that allows the circuit to associate a prediction with the sensory event it precedes. In other words, the anatomy itself encodes the teaching signal.

The second discovery addresses the credit assignment problem, a puzzle that arises whenever learning must be distributed across network stages. If synapses at multiple sites need to change to produce a single learned behavior, how does each site know how much it should contribute? The connectome showed that the wiring between network stages is structured in a way that solves this problem. MG cells, which sit between the granule cell input layer and the output neurons, come in two subtypes, MG+ and MG-, distinguished by the relative positions of their axons and basal dendrites. These subtypes connect selectively to ON and OFF output cells respectively, matching the polarity of the sensory signals they carry. Recurrent connections among MG cells are likewise structured, linking cells of the same subtype rather than connecting indiscriminately, and this recurrence accelerates the generation of the sensory prediction and sharpens its cancellation.

To confirm that this intricate wiring actually performs as its structure suggests, the team built a computational model constrained by electrophysiological recordings from the fish brain. The simulations showed that when the connectivity follows the patterns observed in the connectome, multiple sites of plasticity cooperate to overcome their individual limitations. Plasticity at the granule cell to MG cell synapses does the heavy lifting early in learning, while plasticity at the granule cell to output cell synapses refines the result, and the recurrent MG circuitry amplifies the instructive signals where needed. The outcome is cancellation that is fast, accurate and robust to noise, precisely the properties measured in the living fish. When the researchers perturbed the model, breaking the selectivity of the connections or removing the recurrence, cancellation degraded in ways that matched the specific computational roles assigned to each wiring feature.

The model also illuminated how the circuit handles the continual learning problem without catastrophic forgetting. During periods of unpredictable sensory noise, changes in the excitatory granule cell inputs and the inhibitory MG cell inputs to output cells oppose one another, so that the circuit can absorb disruption without losing its stored predictions. SG cells appear to play a complementary role, enhancing output responses to unpredictable stimuli while MG cells suppress responses to predictable ones, together producing a cleaned-up sensory signal in which behaviorally relevant surprises are preserved. This division of labor across cell types and synaptic sites offers a concrete biological solution to a problem that machine learning researchers have attacked with techniques such as elastic weight consolidation and complementary learning systems.

Beyond its implications for learning theory, the study demonstrates the maturing power of connectomics as a tool for deciphering neural computation. Just as wiring diagrams of the fly central complex and the mushroom body revealed motifs for navigation and associative memory, this reconstruction of a vertebrate cerebellum-like circuit shows how structured connectivity can implement a learning algorithm with known theoretical properties. The authors emphasize that the combination of connectomics, cell-type-specific physiological recordings and computational modeling was essential; none of the three approaches alone could have resolved how distributed plasticity produces fast, accurate and continually updated sensory prediction. The full electron microscopy dataset, along with all analysis code and model files, has been made publicly available through browsable platforms and archived repositories, inviting the wider community to explore the circuit further.

The findings resonate far beyond electric fish. Cerebellum-like structures appear across the vertebrate lineage, and the cerebellum itself is thought to perform analogous predictive computations for motor control, with distributed synaptic plasticity across its granular, molecular and deep nuclear layers. Understanding how a small, tractable circuit coordinates learning at multiple sites provides a template for investigating the same question in larger brains, including our own. As connectomic datasets grow to encompass ever larger volumes of mammalian tissue, the electric fish ELL stands as a proof of principle: given the complete wiring diagram, grounded in physiology and captured in a computational model, the algorithm of a learning circuit can finally be read directly from its wiring.

Subject of Research: Connectomic analysis of sensory prediction and distributed synaptic plasticity in a cerebellum-like electrosensory circuit

Article Title: Connectome analysis of a cerebellum-like circuit for sensory prediction

Article References: Perks, K. E., Petkova, M. D., Muller, S. Z., Genecin, M., Ghatare, A., Schalek, R., Wu, Y., Januszewski, M., Jain, V., Lichtman, J. W., Abbott, L. F., & Sawtell, N. B. (2026). Connectome analysis of a cerebellum-like circuit for sensory prediction. Nature. https://doi.org/10.1038/s41586-026-10690-6

Image Credits: AI Generated

DOI: 10.1038/s41586-026-10690-6

Keywords: connectomics, electric fish, cerebellum-like circuit, sensory prediction, synaptic plasticity, continual learning, electrosensory lobe, credit assignment, corollary discharge, negative imaging, neural circuits, computational modeling

The work targets a long-standing pain point in multi-agent reinforcement learning (MARL), the dominant modern approach to decentralized traffic signal control. In a MARL setup, each intersection hosts an agent that learns its own policy — when to switch phases, how long to hold green, when to yield to cross traffic. The trouble is that the global reward, such as a network-wide reduction in travel time, is shared among all agents. Determining which intersection’s decisions actually produced the improvement is the credit assignment problem, and it is notoriously difficult when the effects of one signal ripple through neighboring blocks.

Conventional solutions tackle credit assignment with critic networks that attempt to decompose the global reward into per-agent contributions. These approaches carry two fundamental limitations, the authors argue. First, they are sample-hungry: agents must engage in extensive trial-and-error interactions with the traffic environment before the critics learn to allocate credit meaningfully, and every trial in a live or simulated network is expensive. Second, the critics must possess strong extrapolation and generalization capabilities — they must interpret policies they were never trained on and still distribute credit sensibly. When traffic patterns shift, as they constantly do, critic networks can produce misleading credit signals that push agents toward worse policies.

BootDiffLights sidesteps this entire machinery. The key insight is a reframing: rather than asking how to divide a global reward among agents, the framework adopts a semi-offline multi-agent reinforcement learning paradigm that transforms the problem from credit assignment into black-box optimization. In practical terms, the system no longer focuses on exploring optimal policies directly. Instead, it concentrates on generating optimal sample data — high-quality trajectories of traffic states and actions — and optimizes in that data space. The generative engine doing this work is a conditional diffusion model, the same class of model that has powered recent breakthroughs in image synthesis and, increasingly, decision-making.

Diffusion models work by learning to reverse a gradual noising process. Trained on data, they learn to transform pure noise into samples that resemble the training distribution, and conditioning signals can steer what they generate. In BootDiffLights, the conditional diffusion model serves as an implicit credit-assignment mechanism: it learns the relationship between traffic conditions, agent behaviors, and resulting rewards, and can then synthesize reward signals consistent with desired regional outcomes. Because the model captures the joint structure of the multi-intersection system, it implicitly encodes how each signal’s actions contribute to network-level performance — without an explicit critic network attempting the decomposition.

The semi-offline aspect matters as much as the diffusion engine. Fully offline reinforcement learning restricts agents to a fixed dataset, which is safe but often suboptimal; fully online learning requires costly and sometimes dangerous exploration. A semi-offline approach lets the system bootstrap from logged traffic data while selectively gathering new interaction data where it matters most. This is what enables BootDiffLights to improve sampling efficiency: the diffusion model can propose promising regions of the behavior space, and the learning system refines its data generation around those regions rather than wandering blindly through trial and error. The framework essentially shifts computational effort from interacting with the environment to generating better synthetic experience.

The authors evaluated the framework on both synthetic and real-world traffic datasets, running all experiments with five random seeds and reporting results as mean plus or minus standard deviation — a rigorous protocol that guards against lucky draws inflating performance claims. Across these benchmarks, BootDiffLights outperformed state-of-the-art baselines on regional traffic metrics, including average queue length and travel time. The comparison field includes well-established MARL techniques built on value decomposition and multi-agent actor-critic architectures, meaning the generative approach did not merely match conventional methods but surpassed them on the metrics that commuters actually feel.

The significance of this work extends beyond one benchmark suite. Traffic signal control has been dominated for decades by fixed-time plans and adaptive schemes such as SCOOT and SCATS, with reinforcement learning offering a data-driven alternative since the late 2010s. The field’s persistent obstacle has been scale: methods that work beautifully on a handful of intersections degrade as networks grow, because credit assignment becomes combinatorially harder. By replacing explicit credit decomposition with generative reward synthesis, BootDiffLights offers a path that may scale more gracefully. The diffusion model’s capacity to model complex joint distributions could prove decisive in dense urban cores where dozens of signals interact within a few blocks.

The approach also connects to a broader trend in machine learning: the migration of generative modeling into sequential decision-making. Diffusion-based policies and planners have recently shown impressive results in robotics and offline control, and BootDiffLights brings that momentum to a civic infrastructure problem with enormous practical stakes. Congestion costs cities billions annually in lost time and fuel, and emissions from idling vehicles degrade urban air quality. A control framework that learns efficiently from existing data, requires no explicit reward decomposition, and improves region-wide metrics could translate into real deployments — though the authors note that data availability constraints limit what can be shared publicly for now.

Challenges remain before such systems direct real intersections. Diffusion models carry their own computational burden at inference time, and traffic agencies will demand robustness guarantees that simulation studies alone cannot fully provide. Yet the conceptual shift is compelling: where previous frameworks asked a critic to guess who deserved credit, BootDiffLights asks a generative model to imagine the rewards worth pursuing — and then optimizes toward that imagination. For a problem as tangled as city-wide traffic coordination, that creative reframing may be exactly the kind of fresh thinking the field has been waiting for.

Subject of Research: Diffusion-based reward generation for decentralized multi-intersection traffic signal control in semi-offline multi-agent reinforcement learning

Article Title: BootDiffLights: Decentralized traffic signal control via diffusion-based reward generation

Article References: Hao, B., He, J., Zheng, J., Dong, W., Yang, Q., & Ma, X. (2026). BootDiffLights: Decentralized traffic signal control via diffusion-based reward generation. Applied Intelligence, 56(15), Article 437. https://doi.org/10.1007/s10489-026-07487-5

Image Credits: AI Generated

DOI: 10.1007/s10489-026-07487-5

Keywords: diffusion model, traffic signal control, multi-agent reinforcement learning, credit assignment, offline reinforcement learning, semi-offline learning, black-box optimization, reward generation, queue length, travel time, smart infrastructure, Applied Intelligence