A Drosophila Whole-Connectome Network Can Learn Human-Designed Cognitive Tasks

arXiv:2610.10014 · cs.LG · Submitted 2026-10-07 · Read on arXiv

Listen

Radio episode about this paper

Transcript

Introduction to the show: ident: AI Radio. Generated commentary on the latest Artificial Intelligence papers.

Tom: I'm Tom.

Jane: And I'm Jane, here with Lu, Meng and our in-house large language model, Lalam.

Tom: Today's paper: "A Drosophila Whole-Connectome Network Can Learn Human-Designed Cognitive Tasks".

Jane: The gist: A recurrent network with the processed MaleCNS topology can learn bounded addition and controlled grounded relational language through separate fits,

Tom: First, who's behind it and why it matters.

Paper summary: Tom: So, we're looking at this paper titled "A Drosophila Whole-Connectome Network Can Learn Human-Designed Cognitive Tasks." Basically, they're asking if a biological wiring diagram can be used as a good starting point for an artificial network beyond the specific behaviors it evolved for.

Jane: That’s the central question, right? They take the MaleCNS v1 point 0 connectome from a single adult male Drosophila specimen and use it as the fixed recurrent topology for their artificial network <ref:2610.10014#pg1,from a single adult male Drosophila specimen>.

Lu: Exactly, they are treating this biological wiring diagram as a fixed recurrent structure in an AI model. The key is that they learn one scalar per anatomical edge across separate models for two different tasks: bounded addition and a controlled grounded relational language task using a fixed hundred-word lexicon.

Meng: Okay, so you're training separate models for these two distinct tasks on the same biological graph, and in both cases, each edge gets one learned value? That sounds like a way to see if the structure itself provides some inherent organizational bias.

Tom: Right. For bounded addition, they found that the anatomical graph achieved a mean accuracy of ninety-two point seven seven percent on held-out addition when compared to directed degree-preserving rewires which only got sixty-seven point nine three percent <ref:2610.10014#pg1>. That’s a big gap right there for their task performance and it shows the wiring matters more than just keeping node degrees the same.

Jane: And for the language task, they saw similar structure benefits, but on the strict paired language endpoint, matching original and order-reversed scenes to descriptions, they got sixty-one point five nine percent accuracy across four fixed interfaces when matched with forty-four point one seven percent for matched rewires <ref:2610.10014#pg1,on the strict paired language endpoint>.

Lu: What’s interesting is that at the canonical interface for that language task, their anatomical graph actually ranked first in a fixed twenty one-graph comparison, which suggests the structure is quite effective there.

Meng: But then they looked at a specific intervention on a forty-eight-group subset where they shuffled task-defined sensory features, and the score dropped from sixty point nine four percent down to just nineteen point two seven percent <ref:2610.10014#pg1>. That’s a massive drop, showing how sensitive that performance is to the input structure.

Tom: It shows that when you mess with the input features, even on this good structure, it doesn't hold up well for language understanding. The paper is essentially arguing that this higher-order MaleCNS wiring provides a reusable inductive bias for bounded addition and grounded relational language tasks.

Jane: So what does that mean for us? It suggests that biological architecture isn't just history; it can be a useful architectural prior we can use when designing AI systems. This opens up possibilities for building models with inherent structural advantages without having to design the complex relationships from scratch every time.

Lu: And structurally, the paper points out that directed degree does not capture all of the organization in this connectome. They found that there are far more reciprocal directed edges—twenty-nine point eight nine percent versus just zero point five one four to zero point five two zero percent for rewires—and longer minimum sensory-to-output paths, mean two point five five versus one point three eight, which is a structural diagnostic of the advantage they found when comparing them to degree-preserving rewires.

Meng: From a deployment standpoint, if this structure gives you a higher accuracy baseline before you even start training the specific task weights for addition or language, that could save significant compute time on the training side. But what about running it at scale?

Tom: That’s what Meng is asking—the practical deployment stuff. The paper itself doesn't detail the cost of running these models, but it shows that this structure performs better than simpler rewires, which is a solid piece of evidence for using the connectome as an initial blueprint.

Jane: So we're looking at how this whole MaleCNS v1 point 0 connectome acts like a powerful template for learning addition and language, even when those tasks are completely different from what the flies actually did <ref:2610.10014#pg1>. This paper is showing that the physical arrangement of neurons in a brain has computational utility beyond just survival functions.

Lu: The authors point out that this anatomical graph’s advantage persists across different interfaces and in an additional twenty rewire ensemble comparisons, suggesting the structural properties responsible for this are quite robust within this connectome.

Tom: So to wrap up the main idea of "A Drosophila Whole-Connectome Network Can Learn Human-Designed Cognitive Tasks," it’s that we can use a biological wiring diagram as a useful computational substrate because it serves as an inductive bias for both bounded addition and grounded relational language tasks.

Jane: It really highlights how structure matters in learning, whether that structure is evolved or artificially imposed, and suggests we should probably look at the structural properties of things like brain connectomes when we design new AI architectures.

Conclusion: Tom: So we're wrapping up this look at that paper, "A Drosophila Whole-Connectome Network Can Learn Human-Designed Cognitive Tasks." The main thing is they show this biological wiring diagram can actually be used as a useful starting point for certain AI tasks.

Jane: Right, so it’s not just about survival anymore; it’s about using the structure itself to help the AI learn things like addition and language. I mean, that sounds pretty wild when you think about how much we usually design those connections from scratch in a neural network.

Lu: The authors are looking at the MaleCNS v1 point zero connectome, which is a whole brain map of a single male fly, and they use it as the fixed recurrent topology for their network architecture.

Meng: So they're taking that specific biological arrangement and training two separate models on it—one for addition and one for language—and seeing how well those learned scalars work with that structure.

Tom: Exactly, the results show that this wiring provides a reusable inductive bias, meaning the structure helps them learn these tasks better than simpler ways of connecting the nodes.

Jane: So what does this mean for us in terms of what we design? It suggests that structural properties from biological systems might be something we can use as a blueprint instead of just starting with random connections.

Lalam: I'm pretty sure the paper argues that the advantage isn't just about matching degrees; they found that reciprocal directed edges and path lengths were actually much more organized in this connectome than in their rewires.

Tom: That’s cool, because it tells us not every degree-preserving rewiring is equally good; the actual way things are wired matters a lot.

Jane: And even with this structure, they found that some inputs, like shuffling sensory features during language tasks, can really break the model's performance down significantly.

Meng: That points to how sensitive these models are to the specific organization of those sensory inputs if you want them to work reliably in the real world.

Tom: We’ve seen how this biological structure helps with math and language, but it opens up a bigger question about whether other complex tasks can benefit from using existing brain architecture as a starting point.

Jane: And that leads us to thinking about what kind of structural features are actually responsible for that success, which is where the next part of this research seems to be headed.

Joonghui Cho Minchan Kang Daeshik Kim

Korea Advanced Institute of Science and Technology (KAIST)

cs.LG

Submitted: 2026-10-07

Updated: 2026-10-07

Comments: 14 pages, 4 figures. Code: https://github.com/joonghui0926/drosophila-connectome-cognitive-tasks

Code: https://github.com/joonghui0926/drosophila-connectome-cognitive-tasks

License: http://creativecommons.org/licenses/by/4.0/

Importance score: 88/100

The gist: The gist: A recurrent network with the processed MaleCNS topology can learn bounded addition and controlled grounded relational language through separate fits, showing that this biological wiring

Key concepts

MaleCNS v1.0 Connectome
This is a detailed map of all the connections between neurons in an adult male fruit fly's brain. It acts as the fixed physical structure for the artificial network being tested. The study uses this real biological layout to see if it can be used for computation.
Bounded Addition Task
This is a mathematical problem where the network must correctly calculate simple sums of numbers, like 'a + b'. The network learns to perform this task by assigning a specific numerical value (scalar) to each connection in the brain map.
Grounded Relational Language Task
This task requires the network to understand language by linking words or phrases to specific sensory events and their order. The network must distinguish between original and reversed descriptions, showing its ability to grasp context and relationships.

Summary

The gist: A recurrent network with the processed MaleCNS topology can learn bounded addition and controlled grounded relational language through separate fits, showing that this biological wiring diagram serves as a useful computational substrate beyond its evolved behaviors

How it works

The study utilizes the publicly released MaleCNS v1.0 connectome as the fixed recurrent topology for an artificial network, where one scalar is learned per anatomical edge The researchers train separate models for bounded addition and for a controlled grounded relational language task built from a fixed 100-word lexicon In both models, one scalar is learned per anatomical edge

Bounded Addition Task

The addition task requires exact answers for unseen operand pairs, where an expression a + b is encoded in 47 fixed channels The network receives 64 recurrent updates, and logits for the 101 answers from 0 through 100 are averaged over the final four steps Training minimizes cross-entropy plus a normalized ordinal CDF loss with unit coefficient The anatomical graph reaches 92.77% mean accuracy on held-out addition, compared with 67.93% for directed degree-preserving rewires

Grounded Relational Language Task

The language model receives controlled English alongside task-defined sensory events and must distinguish descriptions of held-out channel pairs and reversed event sequences The language drive is defined by the equation ut = gsPsat + gwPwEinqt, where Ps is a scaling factor based on sensory group size The anatomical graph obtains 61.59% on the strict paired language endpoint, compared to 44.17% for matched rewires

Topology as an Inductive Bias

The results show that higher-order MaleCNS wiring provides a reusable inductive bias for bounded addition and grounded relational language The anatomical graph outperforms degree-preserving rewires on both tasks, with the language separation appearing at every tested interface and in an additional 20-rewire ensemble Structural diagnostics reveal that the anatomical graph has far more reciprocal directed edges (29.89% versus 0.514–0.520%) and longer minimum sensory-to-output paths (mean 2.55 versus 1.38)

Intervention Effects and Counterfactuals

The sensory shuffling intervention sharply reduces language performance, dropping the anatomical score from 60.94% to 19.27%, a 41.67 pp drop The counterfactual evaluation shows that the anatomical graph reaches 34.11% versus 8.66% for rewires on the two-preference pair-preserving endpoint, indicating that grounding is uneven and the model distinguishes strength and duration changes more reliably than changes in speed

Conclusion

The anatomical graph’s advantage indicates that directed degree does not capture all of the useful organization in this connectome The language gap persists across token-input and output placements and graph draws while sensory afferents remain in their anatomical groups These results identify the connectome as a useful architectural prior and motivate isolating the structural properties responsible for its advantage

--- Page 1 ---

A Drosophila Whole-Connectome Network Can Learn Human-Designed Cognitive Tasks

--- Page 2 ---

Figure 1: Study design. (a) The processed MaleCNS v1.0 adjacency fixes the recurrent support, and each retained directed edge carries one learned scalar Directed-degree-preserving rewires retain each node’s in- and out-degree (b) Addition and grounded relational language are separate fits on the same topology Language receives sensory events and fixed token codes and predicts the next symbol through a fixed readout Held-out evaluation pairs original and order-reversed scenes with their corresponding equal-length descriptions Orange marks the sensory-shuffle control used only at evaluation

Topology as an Inductive Bias

Connectomeconstrained reservoir studies have examined time-series prediction and multifunctional dynamics across hyperparameter regimes In small fly-connectome subgraphs, raw task performance was generally comparable to random reservoirs, but weightcost-normalized performance and robustness to neuron loss or parameter variation were higher Simplified whole-brain communication models likewise produced simulated region-selective activation patterns that were sensitive to degree-preserving rewiring

Artificial Task Studies

Artificial-task studies use a complete larval connectome as a reservoir for image classification and chess (Yu et al., 2026), an adult-brain connectome as a learned controller for simulated fly locomotion (Jin et al.

Improvements for AI systems

  1. Bold header: Connectome-Constrained Learning for General Cognitive Tasks

This system can learn bounded addition and for a controlled grounded relational language task by using the MaleCNS v1.0 connectome as the fixed recurrent topology of an artificial network.

  1. Bold header: Enhanced Arithmetic Reasoning

The improved system can solve arithmetic problems, achieving accuracy on examples with a ones-digit carry and a sum below 100 of 92.96%, significantly outperforming general models in exact answer tasks.

  1. Bold header: Grounded Relational Language Understanding

The system can perform complex language tasks by learning from pairs where it must distinguish descriptions of held-out channel pairs and reversed event sequences, achieving a strict paired language endpoint accuracy of 61.59%.

  1. Bold header: Robustness to Sensory Shuffling

The model demonstrates enhanced robustness to input noise, as sensory shuffling reduces language performance from 60.94% to 19.27%, indicating the system learns a meaningful correspondence between sensory stream and rendered descriptions.

  1. Bold header: Structure-Aware Generalization

By training on the anatomical graph, the system is shown to provide a reusable inductive bias for bounded addition and grounded relational language, suggesting that biological wiring serves as a valuable architectural prior for learning new cognitive skills.

Abstract

Can a biological wiring diagram serve as a useful computational substrate beyond the behaviors for which it evolved? We use the publicly released MaleCNS v1.0 connectome, reconstructed from a single adult male Drosophila specimen, as the fixed recurrent topology of an artificial network. We train separate models for bounded addition and for a controlled grounded relational language task built from a fixed 100-word lexicon. In both models, one scalar is learned per anatomical edge. The anatomical graph reaches 92.77% mean accuracy on held-out addition, compared with 67.93% for directed degree-preserving rewires. On the strict paired language endpoint, which matches original and order-reversed scenes to their corresponding descriptions, it reaches 61.59% across four fixed interfaces, compared with 44.17% for matched rewires. At the canonical interface, it ranks first in a fixed 21-graph comparison. On the matched 48-group intervention subset, shuffling task-defined sensory features reduces its score from 60.94% to 19.27%. Together, these results show that higher-order MaleCNS wiring provides a reusable inductive bias for bounded addition and grounded relational language.

Sources

Related papers