Drosophila Connectome Topology Provides No Measurable Language- Model Advantage: A Preregistered Negative Experimental Program

For decades, a quiet hope has persisted in the world of artificial intelligence: that the messy, biological wiring of a living brain might hold the secret to building smarter machines. The idea is simple. Evolution has spent millions of years refining the connections between neurons in animals, creating networks that are efficient, modular, and surprisingly good at learning. If we could copy those specific connection patterns—the "wiring diagram"—and paste them into a computer program designed to understand language, perhaps the computer would learn faster or think more clearly. This hope is not just a metaphor; it is a testable question. Can the physical map of a brain serve as a helpful starting point, or "inductive bias," for a machine learning model? To find out, researchers turned to the fruit fly, Drosophila melanogaster. Scientists have already mapped the entire nervous system of an adult fly, identifying roughly 139,000 neurons and tens of millions of connections. This map, known as a connectome, is the most detailed wiring diagram of any animal brain available. The question now was whether this specific biological blueprint could actually improve a language computer, or if the complexity of the fly brain was simply too different from the task of processing human words.

A researcher set out to answer this question with a rigorous, pre-planned experiment. They did not simply guess that the fly's brain would help; they designed a series of strict tests where the rules for winning were written down before any data was collected. The goal was to see if using the fly's wiring diagram could make a small language model better at predicting the next word in a sentence, compared to using a random set of connections or no special map at all. They tested several different ways to use the fly's anatomy. First, they tried using the map as a guide for how the computer should pay attention to different parts of a sentence. Then, they replaced a standard part of the computer's processing engine with a frozen, unchangeable copy of the fly's olfactory system, which is famous for expanding information and filtering it. Finally, they tested a more complex idea inspired by how flies store memories: a system where information is hidden and can only be retrieved with a specific "reminder" signal, a concept known as a silent engram.

The results were clear and consistent: the fly's wiring diagram provided no measurable advantage. In the first set of tests, where the researcher used the fly's map to guide the computer's attention, the model performed no better than a model with no map at all. In fact, when they compared the fly-based model to a model using a completely random network of the same size, the random network often performed just as well or better. The researcher found that while the fly's map contained interesting patterns, those patterns did not translate into an ability to understand language syntax or facts any better than a blank slate. One specific test involved taking the raw connections from the fly's brain and using them to predict relationships between words. While the fly map could sometimes beat a scrambled version of itself, it failed to beat the baseline of having no map at all. The difference in performance was so small that it was statistically indistinguishable from random chance.

The experiment went further, testing the fly's brain as a replacement for a standard computer component. The researcher took the part of the fly brain that expands sensory input and filters it, and they froze it in place to act as a fixed processor for the language model. They expected this biological structure to be a superior way of handling information. Instead, the computer performed worse than if they had used a simple, random network of the same size. The biological expansion was not a magic key to better language processing; it was, in this context, a less efficient tool than a standard, unstructured network. The researcher also tested the idea of the "silent engram." In flies, memories can be stored in a way that makes them invisible to normal checks, becoming accessible only when a specific chemical signal acts as a reminder. The researcher built a computer system that mimicked this behavior, trying to see if this biological trick for hiding and retrieving memories offered any advantage over standard computer memory. They found that while they could build a system that worked like the fly's memory, a much simpler, generic computer program could do the exact same job just as well. The fly's specific method of hiding memories was not unique or superior; a standard, constrained selector could achieve the same results without needing the complex biological blueprint.

Even the researcher's attempt to use the fly's brain as a tool for finding information, similar to how a fly uses its brain to recognize smells, failed to outperform standard methods. When they used the fly's expansion mechanism as a way to look up data, a random expansion method worked better. The dense, standard computer methods were far superior at retrieving the correct information. The study also looked at the fly's central complex, a part of the brain involved in navigation and holding a sense of direction. They tried to see if this structure could act as a stable memory loop in a computer, holding onto a piece of information over time. The test showed that the fly's structure did not hold the information any better than random networks; in fact, the information faded faster than in the control groups.

Throughout the study, the researcher was careful to distinguish between what they tested and what they did not. They did not claim that the fruit fly's brain is useless for all machine learning, nor did they say that brains are bad models for computers. They simply showed that these specific attempts to copy the fly's wiring into a language model did not work. The 800-million-parameter model they used as a reference point in the background was not proven to be better because of the fly's map; its performance was measured separately and was not part of the connectome experiment. The study concluded that while it is possible to build a computer that uses sparse, family-level communication similar to the fly, the specific topology of the Drosophila projectome offers no advantage over random or learned connections. The biological blueprint, in this case, was not a shortcut to a smarter machine. The experiment demonstrated that evolution's solution for a fly's brain does not automatically solve the problem of teaching a computer to understand human language. The hope that copying a brain's wiring diagram would instantly improve artificial intelligence was, in this specific and carefully measured instance, not supported by the evidence.