A single dendritic compartment that learns without touching a synapse

What the work claims

In-context learning is the ability of a trained sequence model to solve a new task from a few labelled examples in its input, with no weight update. Recent theory has traced this ability, across Transformers and other architectures, to an implicit gradient-descent process running inside the forward pass. The authors point out that spiking neural networks, the substrate closest to biology and the target of neuromorphic chips, have conspicuously failed at this: standard spiking baselines collapse on the Garg-2022 function-learning benchmark once the task dimension climbs past about twenty. Their claim is that this failure is structural and misdiagnosed. Prior designs treated the dendrite as a passive conduit and routed adaptation through synaptic plasticity at inference. Instead, they argue, the subthreshold dynamics of a single dendritic compartment already are a complete online learning algorithm.¹

The concrete deliverable is DendriCL, a single-layer compartmental spiking network whose apical compartment update is, by construction, mathematically identical to leaky online Widrow-Hoff least-mean-squares, an adaptive-filter rule dating to 1960. This is a method-and-mechanism paper backed by simulation, not a hardware result and not a biological measurement. Its boldness is the inversion it proposes: that depth, attention and inference-time plasticity, the three ingredients usually assumed necessary for general in-context learning, are all dispensable, and that one compartment with the right dynamics suffices.

How it works

The model is loosely shaped after a layer-5 cortical pyramidal neuron, which has an apical dendrite receiving top-down input, basal dendrites receiving bottom-up input, and a soma that integrates both. In DendriCL the apical compartment holds a multidimensional subthreshold membrane potential that is not reset when the neuron spikes, so it can carry information across the whole sequence of examples. The authors show that the recurrence governing this potential has exactly the form of a leaky least-mean-squares update: each new labelled example nudges the membrane state toward the task's underlying parameter vector, the same way an adaptive filter incrementally corrects its estimate. Crucially, the synaptic weights that define this dynamic are trained once, end to end, and then frozen. Learning at inference happens in the evolving voltage, not in any changing connection.

The evidence that this is what is happening, rather than a black box that merely performs well, is a linear probe: a simple linear decoder recovers the reference online-LMS trajectory directly from the apical membrane potential at an R-squared of 0.93. That is the paper's strongest internal result, because it shows the algorithm is structurally embedded in the dynamics rather than approximated somewhere diffuse. On the benchmark itself, DendriCL stays stable across task dimensions from five to fifty, where Spikformer manages an R-squared of only 0.72 at dimension twenty, a plain leaky-integrate-and-fire network sits near 0.09, and an adaptive-threshold spiking network near 0.01. It is also the only model that remains seed-stable past dimension thirty, a regime where even dense Transformers become bimodal and unreliable and fail beyond dimension forty. The authors further report roughly a fourfold reduction in spike count versus the plain spiking baseline and a projected order-of-magnitude energy advantage on Loihi-class hardware.

Where a skeptic should push

The most load-bearing assumption is that the clean apical-LMS equivalence, proved for the compartment in isolation, survives contact with the full spiking neuron. The authors are candid that their theorem does not formally handle the leaky-integrate-and-fire reset nonlinearity at the soma; they lean on the empirical probe correspondence to argue it holds in practice. That is a reasonable defence, but it means the headline mathematical identity is exact only for a linearised piece of the model, and the R-squared of 0.93, while high, is not unity. The equivalence is strong evidence, not a closed proof for the system as run.

The breadth of the evidence is the other place to push. The entire demonstration lives on Garg-2022 synthetic function classes, chiefly linear regression tasks; there is no language, no vision, no real sensory stream. In-context learning of a linear map is a genuine capability, but it is the friendliest possible case for a least-mean-squares mechanism, so the result may partly reflect a match between the toy task and the chosen dynamics. The energy numbers are spike-count instrumentation plus Loihi projections, not measured joules on a chip. And the Transformer failure it contrasts against is estimated from six seeds at one dimension, a small sample for a distributional claim. None of this undermines the core finding, but it keeps the paper firmly in the well-controlled, in-silico, proof-of-principle category.

What organoids might hide below the spike

Organoid intelligence reads its substrate almost entirely through spikes on a microelectrode array, and it usually assumes that to make tissue learn you must change its synapses. This paper is a pointed challenge to both habits, and it does not have to be biologically true to bite, because it is a claim about what a dendritic compartment can compute. Its first implication is about the readout. If a meaningful chunk of a neuron's online computation is carried in a slow, subthreshold, non-reset dendritic voltage, then the spike train an electrode records is a lossy, downstream projection of it. A field that certifies and decodes its computers from spikes alone could be watching the shadow while the arithmetic happens in the membrane, and microelectrode arrays record extracellular spikes while remaining largely blind to subthreshold membrane dynamics by construction, independently of whether this particular mechanism turns out to be biological. Stated as a conditional: if a comparable mechanism ever operated in wet tissue, a spike-only readout would be a lossy projection of it.

The second implication undermines a common design premise. The working assumption that online adaptation requires synaptic plasticity is exactly what DendriCL discards: here adaptation is a property of frozen-weight dynamics, and no synapse changes at inference at all. The honest caveat is that the plasticity has not vanished so much as moved. The weights that shape the apical recurrence were set once by backpropagation-through-time, so the configuration effort is front-loaded into a non-biological training run rather than removed, and this is a silicon existence proof, not a statement about living substrates. Real cortical neurons do have elaborate active dendrites; human layer 2/3 cells have been shown to support dendritic action potentials and local nonlinear computation.² Those dendrites supply the substrate for rich subthreshold nonlinearity, but whether biological dendrites carry the specific time constants an LMS-like estimator needs is unknown, and DendriCL obtained them only through training. So the honest reading is a motivation to look for such a mechanism in tissue, not a capability tissue is shown to hold.

The threat is a hype-correction, and it is stern. DendriCL only works because its weights were tuned end to end by backpropagation-through-time to make the dendritic dynamics implement LMS. The compartment does not spontaneously compute a learning algorithm; it does so after training has sculpted the parameters that shape its subthreshold recurrence. For a grown organoid to do the analogous thing, development or evolution would have to have already set the equivalent constants, and nothing here shows that untrained tissue possesses them. So the paper is not evidence that organoids already run online-LMS in their dendrites. It is evidence that a single compartment can, in principle, host such an algorithm, which reframes what would be worth looking for, and worth interfacing with, rather than certifying that the substrate is secretly doing it today.

The bottom line

Established in simulation: a single-layer compartmental spiking network whose apical dynamics are constructed to equal leaky online least-mean-squares solves a synthetic in-context learning benchmark across task dimensions five to fifty, stays stable where spiking baselines and even Transformers fail, and lets a linear probe recover the learning trajectory from its membrane at R-squared 0.93. Unestablished: that the mechanism generalises beyond linear synthetic tasks, that the projected energy advantage holds when measured on real silicon, and above all that biological neurons actually implement anything like it. For organoid intelligence the durable lesson is a reframing, not a discovery: the computationally interesting state may be subthreshold and dendritic rather than in the spikes we record, and online learning need not mean synaptic change. What would confirm the biological version is an in-vivo or in-tissue recording showing an apical-LMS-like trajectory during a real in-context task; what would deflate it is evidence that the effect is a linear-regression artefact that does not survive richer tasks or the full spiking nonlinearity.