A Turing-complete symbolic computer made of nothing but spikes
Three researchers report the first complete implementation of beta-reduction in the untyped lambda-calculus, the minimal Turing-complete symbolic formalism, executed entirely in the spike dynamics of leaky integrate-and-fire neurons. Every reduction step maps to a named, experimentally documented neural mechanism, every synaptic weight is derived analytically from measured excitatory postsynaptic potential amplitudes, and the closed-loop system reproduces a symbolic oracle on all ten validation classes. The result is a simulation, not a culture. That distinction is where the interesting questions for organoid intelligence live.
Source: Fully Spiking Neural Implementation of Lambda-Calculus Beta-Reduction via Biologically Documented Mechanisms, Lazovskiy, Kazantsev and Stasenko, arXiv:2611.23182v1 [q-bio.NC], submitted 17 November 2026. Primary source. Read: full author HTML, including methods, the parameter derivation and calibration summary, and the limitations discussion, verified inline against every number quoted here. Preprint, not peer reviewed.
What the work claims
Know the genre. This is a computational modelling paper: no biological experiment, no hardware, no organoid, no microelectrode array. Its claim is a constructive existence proof with an unusually strong form of biological grounding. The authors implement beta-reduction, the single rewrite rule of lambda-calculus and the engine of every functional programming language, so that all of the work is done inside a spiking neural simulation: redex detection, subterm segmentation, variable binding detection, De Bruijn index arithmetic, token copying, and maintaining roles between reduction steps.1 No symbolic computation happens outside the spike simulation.
What makes the claim more than another universality demonstration is the mapping. Each computational primitive is attached to a specific mechanism documented in experimental neuroscience: Jeffress-type delay-line coincidence detection finds reducible expressions; a line-attractor balance counter, in the style of oculomotor neural integrators, segments subterms; theta-gamma cross-frequency coupling handles variable binding, in the Lisman-Jensen tradition of hippocampal working memory; a ring of eight neurons carrying a traveling phase wave does depth-index arithmetic; theta-cycle spike-timing dependent plasticity copies tokens; and recurrent persistent firing, modelled on entorhinal and prefrontal delay activity, holds roles between steps.
The validation is closed-loop and quantitative: 100 percent agreement with a symbolic oracle across ten validation classes, covering Church numeral arithmetic (plus, mult, pred), Boolean logic with divergent branches, individual Y-combinator steps, terms with up to six bound-variable occurrences, De Bruijn indices up to 12, noise robustness, and determinism. Each token is represented by 64 neurons in a phase-addressed cube of eight phase channels by eight channels per position.
How it works
The engineering discipline is the real contribution. All neurons follow a current-based leaky integrate-and-fire model with explicit membrane dynamics, exponential synaptic currents, stored axonal delays in a circular ring buffer, and pair-based spike-timing dependent plasticity. The membrane time constants are heterogeneous and justified cell by cell: about 5 to 20 milliseconds for pyramidal cells and fast-spiking interneurons, 2 milliseconds for the fast gate neurons that act as coincidence detectors, and 40 milliseconds for the slow integrators that accumulate coincidences within a theta window. A fast AMPA-like synaptic time constant of 1 millisecond serves coincidence detection.
Crucially, the weights are not trained and not tuned by hand. The authors derive every synaptic weight analytically from measured excitatory postsynaptic potential amplitudes and summarize the derivation and calibration in a dedicated section. Inter-module gates have a narrow operating window, between 0.79 and 0.87 in their gating variable: below it the scan cannot trigger, above it it fires spontaneously. Persistent-activity neurons are quenched by a halt pulse after each reduction step. The whole system runs in a custom Python and NumPy engine with no external simulator.
The cost structure is stated honestly. Nested substitutions allocate gate neurons proportional to the number of bound variables times the argument length, which is quadratic in the worst case, and the authors flag temporal scaling as a limitation. Their proposed next step is deployment on Intel Loihi 2, which they argue would run all populations in parallel at full real-time speed, potentially cutting wall-clock time per reduction step below 100 milliseconds.
Where a skeptic should push
The load-bearing assumption is that the named biological mechanisms survive translation into idealized dynamics. The paper's delays are stored exactly in a ring buffer; real axons jitter. The plasticity window is pair-based and theta-timed; real cortical spike-timing dependent plasticity is triplet- and rate-sensitive, and its effective window drifts with neuromodulation. The phase cube uses exactly eight channels; nothing in biology guarantees such tidy quantization. The authors do test noise robustness and determinism as explicit validation classes, which is more than most simulation papers manage, but the noise model is their own.
Separate demonstrated from asserted. Demonstrated, in simulation: each reduction step type executes with correctly derived weights, and the closed loop matches the oracle on every class tested, which is a real integration result because the classes exercise the full pipeline rather than isolated modules. Asserted: that this constitutes a general neural basis for symbolic reasoning, and the framing that the mechanisms are "biologically documented." The citations behind that framing are genuine and appropriate (the Jeffress model, dendritic coincidence detection in cortical pyramidal cells, entorhinal persistent activity), but "documented in some biological context" is not the same as "assembled this way in any brain." No animal performs lambda reduction.
The scale is also worth dwelling on. Ten validation classes with indices up to 12 and six bound-variable occurrences is a correctness demonstration, not a performance claim. The authors themselves list the sharp edges: free variables are a known encoding limitation, evaluation order is leftmost-outermost only, eta-reduction is absent by design, and nested substitution costs quadratic neuron growth. Universality here is the cheap kind, the kind Wolfram-style systems give away for free: any unbounded-resource claim about Turing completeness sits on top of a finite network that implements only bounded terms. And the hardware result does not exist yet. Loihi 2 deployment is a future-work paragraph, not a measurement.
What Turing-complete spike code asks of organoids
The obvious reading is that this is irrelevant to organoid intelligence because it is pure simulation. I think the opposite: it is one of the most useful specification documents the field has had, precisely because it is honest about what structured spiking computation requires.
Here is the non-obvious implication. The organoid computing programme mostly proposes to train tissue end to end: present inputs, reward desired outputs, let plasticity find a circuit. This paper shows what the alternative looks like. Nothing in the lambda implementation is learned. The network computes a symbolic, verifiable procedure because its weights were derived, not optimized, which means its output can be checked step by step against an oracle. That property, formal correctness, is precisely what gradient-trained biological systems cannot currently offer, and it is what any safety case for computation on living human tissue will eventually demand. The parts list the paper assembles, delay-line coincidence, phase rings, line attractors, theta-coupled windows, persistent activity, reads like a catalogue of dynamical primitives that are individually observable in cultured networks. Several of them, including theta-nested gamma and persistent delay activity, have been reported in cortical organoid electrophysiology. So the paper quietly defines a research direction: instead of asking tissue to learn a task, ask whether a culture can be steered into hosting one of these primitives reliably enough to chain two of them. That is a much smaller, much more testable claim than general biological computing, and this paper provides the acceptance criteria.
Now the threat, and it is a serious one. The paper demonstrates that everything claimed for spike-based biological computing can be done in NumPy on a laptop, with derived weights, exact delays, and reproducible runs. The authors note that Loihi 2 would run the whole system in real time. If universality and even structured symbolic computation are available on silicon with zero incubator cost, zero donor-cell ethics surface, and full observability of every weight, then the biological substrate must justify itself on efficiency, adaptability, or sensing integration, not on computational power. This paper raises that bar by showing how cheap the silicon version of "brain-like computation" already is.
There is a second, subtler warning embedded in the mechanism mapping. Every primitive in the parts list requires specified wiring: exact axonal delays of 4 milliseconds in the fine ring, inhibitory halt pulses of weight minus 15, gate thresholds tuned to a 0.08-wide operating window. Organoids are self-wired; connection topology is an uncontrolled output of development, and axonal delay distributions are whatever they turn out to be. The paper is therefore also an inventory of what "programming a culture" would need to control, and the list is long and quantitative. Anyone claiming an organoid performed symbolic computation should be asked which of these primitives they demonstrated, with what tolerances. That question, made concrete, is this paper's gift to the field's governance discourse.
The bottom line
Established, within the simulation: a complete, mechanistically grounded implementation of lambda-calculus beta-reduction in leaky integrate-and-fire dynamics, with analytically derived weights and 100 percent oracle agreement across ten validation classes including noise robustness and determinism. This is a genuine integration achievement and the best-argued case to date that named experimental mechanisms can be composed into symbolic computation.
Hypothesis, not result: that any of this occurs in real neural tissue, that the noise, plasticity, and wiring tolerances of biological substrates can host these primitives at the required precision, and that a hardware implementation will preserve the correctness. The Loihi 2 experiment, when someone runs it, is the natural next gate.
For organoid intelligence specifically, the actionable test is smaller than the headline: pick one primitive from the parts list, the delay-line coincidence detector is the most electrode-accessible candidate, and demonstrate in a cortical organoid on a microelectrode array that evoked delay-tuned coincidence responses exist, are stable for hours, and can be switched by inhibition. A culture that passes that test has earned the right to be called a computational substrate. One that cannot is not failing Turing completeness; it is failing the engineering prerequisite for everything this paper shows is possible.
Frequently asked questions
Is this done with real neurons or organoids?
No. It is a computer simulation in a custom Python and NumPy engine, with no external simulator, no biological experiment, and no neuromorphic hardware run. Deployment on Intel Loihi 2 is proposed as future work. Any application to living tissue discussed here is our extrapolation, not the authors' claim.
What is lambda-calculus beta-reduction?
Lambda-calculus is a minimal formal system for expressing computation through function abstraction and application. Beta-reduction is its single rewrite rule: applying a function to an argument substitutes the argument into the function's body. It is the theoretical foundation of functional programming languages and is Turing-complete, so a machine that performs beta-reduction can in principle compute anything computable.
What does "biologically documented mechanisms" mean here?
Each circuit component is mapped to a mechanism with experimental support in neuroscience: Jeffress delay-line coincidence detection from binaural hearing, line-attractor integration from the oculomotor system, theta-gamma coupling from hippocampal working-memory models, spike-timing dependent plasticity from cultured neuron studies, and persistent firing from entorhinal and prefrontal recordings. Documented individually, however, does not mean ever assembled in this configuration in a real brain.
Were the synaptic weights trained?
No, and that is a central design claim. All synaptic weights were derived analytically from measured excitatory postsynaptic potential amplitudes, with no hand tuning. The authors include a dedicated parameter derivation and calibration summary section. This is why the system can be checked against a symbolic oracle: its behavior is engineered, not optimized into existence.
How strong is the evidence?
Strong for correctness within the model: closed-loop validation against a symbolic oracle across ten classes, including noise robustness and determinism checks, is unusually rigorous for this genre. Limited for scale: terms are small (indices up to 12, at most six bound-variable occurrences per term), evaluation order is fixed to leftmost-outermost, free variables are unsupported, and nested substitution costs quadratic growth in neurons. It is a preprint and has not been peer reviewed.
Why does this matter for organoid intelligence?
It supplies two things the field lacks: a concrete parts list of dynamical primitives that biological substrates could in principle host, each with tolerances and acceptance criteria, and a demonstration that silicon reproduces spike-based computation cheaply and verifiably. Together they redefine what an organoid must prove: not vague "brain-like computing," but the reliable hosting of specific, named mechanisms, at efficiencies and adaptabilities silicon cannot match.
References
- Lazovskiy, V.A., Kazantsev, V.B. and Stasenko, S.V. Fully Spiking Neural Implementation of Lambda-Calculus Beta-Reduction via Biologically Documented Mechanisms. arXiv preprint arXiv:2611.23182v1 [q-bio.NC]. 2026. https://arxiv.org/abs/2611.23182. Accessed 2026-11-22.