Searching for architectures silicon would never design

What the work claims

This is a primary methods-and-results paper. The central claim has three layers. First, that architecture rankings do not transfer between real-valued and spiking substrates: of 106 architecture programs adapted from the ASI-Arch release, 97 yielded complete paired results, and the rank correlation between the ANN and SNN implementations on WikiText-2 is a Kendall tau of only 0.2997. Worse, on those same identifiers the cross-task correlation between WikiText-2 perplexity and ListOps accuracy collapses from 0.3031 for ANNs to 0.0516 for SNNs, meaning an architecture that spikes well on one task tells you almost nothing about how it will spike on another.¹

Second, that an LLM can search an open program space for native spiking architectures, code, not a fixed configuration graph, subject to interface, causality, and spiking-projection constraints, and that a three-view candidate characterization, the code itself, its stated design rationale, and a behavioral fingerprint of pre-training architecture statistics and network responses, is enough to estimate novelty and to predict performance well enough to spend expensive training only on promising candidates. Third, that the search finds designs worth keeping: over 10 outer iterations with a batch limit of 16 and a budget of 160 evaluations, 151 new architectures were trained at an estimated cost of 132 V100 GPU-days, and three representatives of the final Pareto front, NeuroGate, HomeoResSSM, and LoopMem, each embody distinct mechanisms such as spike-activity-dependent control of recurrent state updates and residual pathways.

The headline numbers: NeuroGate attains 26.4 WikiText-103 perplexity against 27.5 for the strongest ANN expert, DeltaNet, trained under the authors' common protocol, while firing about half as often as the SpikingDeltaNet expert. Estimated architecture-level arithmetic energy falls by 17 to 51 times relative to a dense Transformer of the same size, with LoopMem at the top end and NeuroGate at 31.7 times. On the six Long Range Arena tasks, NeuroGate holds the highest average accuracy among all listed SNNs.

How it works

The pipeline is quality-diversity search where the genome is Python. An initial archive holds 20 expert SNN adaptations of linear recurrent and state-space models. Each iteration, Gemini-2.5-Flash, chosen after a comparison of candidate LLMs, generates or mutates executable architecture code for an SNN block inside a fixed outer model structure. Because raw code edits are noisy, candidates are characterized before any training through three views: the code, the design rationale the LLM writes for it, and a behavioral fingerprint that combines architecture statistics with measured network responses taken before weight training. A surrogate regressor predicts task metrics from the fingerprint, and selection is explicitly two-objective, predicted performance against estimated novelty relative to everything already in the archive, so the search keeps architecturally distinct candidates rather than collapsing onto one family.

The spiking-projection constraint is the subtle engineering choice. Every module must express its input-output map in spikes, but internal state may be continuous, which is what allows discovered mechanisms like NeuroGate's to exist at all: spike activity itself modulates the recurrent update and the output gating pathway, a feedback loop that has no clean analogue in a real-valued network. Energy accounting follows the standard proxy: spike-driven projections count as accumulate operations, continuous computation as multiply-accumulates, each model's own measured firing rate, and a shared wrapper model so comparisons are apples-to-apples. The authors are candid that this is an architecture-level arithmetic estimate, not a measured silicon number.

Where a skeptic should push

The most load-bearing assumption is that surrogate-predicted performance and novelty are faithful enough to guide real training. The paper supports this with offline studies on 223 WikiText-2 and 54 ListOps architectures, and the search ablations show the full multi-island surrogate configuration beats single-island and surrogate-free variants, but the main result is one search run. Smaller-scale studies were repeated three times; the flagship run was not. In an LLM-driven search, run-to-run variance is exactly where interesting-sounding but lucky architectures hide, so the discovered trio should be read as existence proofs of a search method, not as three validated designs.

Second, "native" is native relative to a wrapper. The search varied the internal SNN block but the outer structure, including the spiking feed-forward network and output head of the shared wrapper, was fixed, and those shared components account for most spike-driven operations. The 17 to 51 times energy figure is therefore architecture-level and proxy-based: it assumes ideal accumulate-for-spike hardware, counts arithmetic only, and excludes training, memory movement, and the LLM search itself, which cost 132 V100 GPU-days before any downstream retraining. A 50-times arithmetic reduction at the block level routinely compresses to a small multiple at system level, and the paper never measures a deployed chip.

Third, the comparative frame is narrow. WikiText-103 language modeling and the Long Range Arena suite are the only full-scale benchmarks; there is no speech, no control, no continual-learning test where spiking state dynamics would be expected to shine. And the headline ANN comparison is against DeltaNet, strong but one family, not against the current best dense or hybrid models. None of this makes the core finding, that ANN rankings fail to transfer into spiking computation, any less real; that part rests on 97 paired architectures and is the paper's most solid contribution.

What native architecture means for living tissue

The non-obvious implication for organoid intelligence is that the paper's central warning scales up, hard. Spiking silicon differs from ANN silicon in exactly one respect, binary event-driven communication, and even there the transfer failure is severe: rank correlation of 0.30, cross-task correlation near zero. A cortical organoid differs from a spiking chip in a hundred respects: conductance-based synapses, dozens of cell types, glial modulation, slow homeostatic plasticity, developmental drift, and no compiler. If ANN-derived architectures underuse spiking silicon, then ANN-derived readout and stimulation architectures are almost certainly underusing living tissue by a far wider margin. The field's default stack, rate-code readout layers trained with ridge regression or backprop-through-a-model, is the spiking-field's "ANN adaptation" problem in a wetter form, and nobody has yet run the organoid equivalent of the 97-pair experiment to measure how badly.

The constructive opportunity is that OpenArchEvo hands the field a template for fixing this. Its three-view fingerprint, code, rationale, and pre-training behavior, has a direct wetware translation: characterize a culture cheaply with evoked-response maps, spike-train statistics, and perturbation recovery curves before committing to expensive closed-loop training; use those fingerprints to train a surrogate that predicts task performance per culture; and select stimulation and recording configurations to test as a two-objective problem against novelty, keeping an archive of diverse protocols instead of re-deriving one protocol per lab. Closed-loop organoid experiments currently burn weeks of culture lifetime testing configurations a surrogate could have ruled out, and culture-to-culture variability is the exact setting where novelty-preserving search beats single-track optimization.

The threat deserves equal precision. If open-code search keeps producing spiking-native silicon designs with 50-fold arithmetic-energy advantages that beat ANN baselines at language modeling, one pillar of the organoid pitch, "only living tissue gives you event-efficient temporal computation," keeps eroding on silicon's terms. And the paper prices the meta-level honestly: 132 V100 GPU-days to find three blocks. Any honest accounting of organoid computing must now price in the search overhead of discovering what a given culture can compute, because a substrate whose capabilities must be re-characterized every few weeks of developmental drift carries a recurring search cost that a compiled chip does not. The durable niche for living tissue after this paper is not efficiency, and not even dynamics richness alone; it is that tissue rewrites its own architecture between experiments, and the right response to that is the paper's own lesson, search the harness per substrate, never assume the harness transfers.

The bottom line

Established: ANN and SNN implementations of the same 97 architectures rank only weakly alike (Kendall tau 0.2997), and SNN cross-task transfer of rankings nearly vanishes (0.0516); an LLM-evolved open-code search with fingerprint-based surrogate selection finds spiking-native architectures, of which NeuroGate reaches 26.4 WikiText-103 perplexity versus 27.5 for ANN DeltaNet at roughly half the expert's firing rate, with estimated arithmetic-energy reductions of 17 to 51 times over a same-size dense Transformer. Hypothesis, not established: that the discovered trio are robustly better rather than one-run fortunate; that architecture-level arithmetic energy survives translation into measured system energy; and that fingerprint-guided search extrapolates beyond these two benchmark families. What would confirm the program is a repeated full-scale search with per-run Pareto-front overlap analysis and a hardware deployment; what would break the "native advantage" claim is a repeated search in which surrogate-guided evolution fails to beat well-tuned spiking adaptations of strong ANN designs.