C Elegans Connectome Reservoir Benchmarks Show Randomized Wiring Often Beats the Biological Brain
Felix S. Reimers, Ola Huse Ramstad, Aliaksandr Hubin, and Stefano Nichele of Oslo Metropolitan University posted “Benchmarking the Connectomes of Caenorhabditis elegans within the Reservoir Computing Framework” to arXiv on September 24, 2026 (arXiv:2609.30508). The question they test is the founding question of whole brain emulation: does the wiring, by itself, carry the computation? Their answer, from the only organism whose complete nervous-system connectome has been published at multiple ages and under multiple measurement methods, is that on the tasks they ran, biological wiring and bio-informed input-output placement did not reliably beat equal-sized randomized null reservoirs.
The direct answer to the consciousness question this raises is a boundary, not a capability. A connectome graph, dropped into a reservoir-computing testbed with no plasticity and minimal preprocessing, performs as a generic recurrent substrate whose task performance depends more on configuration than on biological provenance. That is evidence about what structure alone supplies, and it constrains every claim that copies anatomy far enough to copy function. A connectome is where emulation starts, and this paper measures how far it gets on its own.
The experiment
C. elegans is the first organism for which physical connectomes covering the whole nervous system were published. The paper uses connectomes derived at different ages of the animal and based on three different ways of measuring intercellular connections. With minimal preprocessing each connectome becomes an echo state network, a recurrent reservoir whose internal weights stay fixed while a small readout module trains on the reservoir’s outputs. Training and testing ran across neuro-inspired tasks, with randomized null models of equal size run identically for comparison, and the whole procedure repeated across reservoir configurations.
| Factor varied | Range | Effect observed |
|---|---|---|
| Connectome age | Different developmental ages | Outcomes vary, no clear trend |
| Wiring measure | Three intercellular measurement methods | Results depend on derivation |
| Input/output placement | Bio-informed vs other configurations | Bio-informed placement not reliably better |
| Wiring itself | Biological vs randomized nulls | Nulls often outperform |
The headline finding is the last row. The randomized null models often outperformed the original connectomes on the chosen benchmarks. At the same time the results depend heavily on reservoir configuration and on how the connectome was derived, so the paper’s own summary is that biological wiring does not necessarily lead to better performance under this testbed.
What a negative result of this kind establishes
A null result here is informative in a way a positive result would not be. If biological wiring had decisively beaten random graphs on every task, structure-first emulation would have its cleanest confirmation. Instead the result lands where the field’s careful statements have long sat. A connectome is a set of constraints on dynamics, and the dynamics are what compute. The Sandberg and Bostrom whole brain emulation roadmap treated simulation parameters as a bottleneck equal to structural acquisition for exactly this reason, and the OpenWorm project documented the missing mechanisms that appear the moment a wiring diagram is asked to generate behavior rather than benchmark scores.
The paper’s age-and-method sensitivity sharpens the point. Connectomes derived at different ages produced varying outcomes with no clear trend, which says the derivation choices are not noise to be averaged away but parameters that change what the graph computes. The MaleCNS and FlyWire comparison makes the same observation at fly scale, where synapse counts supply anatomical weights while physiological dynamics require separate evidence. At 302 neurons the worm makes the lesson cheap to run. At 140,000 neurons it becomes a program, as in the FlyCNS embodied controller and MaleCNS spectral mixing analyses, where connectome structure supplied an effective communication prior only once the sensorimotor loop was closed through a body.
The companion framework, structure to function with a measured bridge
The companion preprint, “Connectome-to-Function: Conditional Generative Latent Representations for Reservoir Computing” (arXiv:2609.06093, Zhuolin Yu, Xingyu Liu, Yuanhao Jia, Yunhang Xiao, Hairuo Xue, Feihan Sun, and Guozhang Chen, posted September 5, 2026), builds the machinery the benchmark result demands. It encodes connectome graphs into a compact conditional generative latent space, reconstructs observed connectivity with a mean edge-reconstruction AUC up to 0.910, and generates candidate connectomes from the same space. When connectome-derived graphs run as reservoir substrates, the latent space captures functional variation across reservoir-computing experiments with cross-validated R-squared up to approximately 0.87.
The interpretability analysis is the part that reaches toward circuit function. Memory performance associated with reciprocal recurrent connectivity, while prediction and classification associated more strongly with spectral properties of the recurrent network. Those are task-specific structure-function mappings, learned rather than assumed, and they are the kind of quantitative bridge that turns a wiring diagram into a statement about computation.
Comparison to The Consciousness AI
The Consciousness AI project keeps structural wiring, dynamic parameters, and validation as distinct layers, under functionalist emergentism where consciousness is an emergent property of organized dynamics. The worm reservoir benchmark is a direct empirical support for that separation. If the biological graph itself does not determine task computation in a fixed-weight testbed, then an emulation architecture that treats the connectome as its dynamic model, rather than as one input to a parameterized dynamic model, inherits a measurable error. The project’s Layer 1 LIF substrate takes dynamics, not anatomy, as the thing to be organized, and this result is one reason that choice is defensible.
The negative result also sets a test the project’s consoles can state precisely. For any substrate, adding biological provenance to a recurrent network should buy measurable functional structure, or the provenance is decorative. The connectome-to-function latent space is exactly the instrument that would run that comparison on real connectomes, and no equivalent instrument exists yet in the project’s own stack. Stating that openly is the honest position, and the brain emulation section tracks the field’s progress toward it.
What the benchmarks leave open
Three limits bound the result. The tasks are neuro-inspired rather than the worm’s own behavioral repertoire, so the nulls outperforming biology says nothing about chemotaxis or locomotion. The reservoir-computing testbed fixes the reservoir weights, which removes the plasticity that makes real neural wiring adaptive, and a connectome evolved for learning may show its value only under learning rules. And the worm’s connectome is measured with substantial uncertainty, which the three derivation methods in the paper make visible.
What survives those limits is the paper’s framing contribution. The connectome is now a benchmarkable computational object with a known failure mode, structure without configuration does not compute distinctively, and the field has a testbed to iterate on. Whether structure plus dynamics ever suffices for function, and whether function at worm scale extends to anything worth calling experience, is the question the flagship overview of AI consciousness research tracks across the field, and the smallest complete nervous system remains its cheapest testbed.