Neuroscience Keeps Solving the Same Problems Twice

Introduction

In 1986, Apostolos Georgopoulos and his colleagues published a paper in Science [2] that became one of the most influential results in motor neuroscience. They recorded from neurons in motor cortex while monkeys reached in different directions, and found that each neuron was "tuned" to a preferred direction. Add up these preferences across a population, weighted by activity, and you get a "population vector" pointing where the monkey is about to reach. The result was clean, intuitive, and felt like it answered a basic question: what does motor cortex represent? Movement direction.

Fourteen years later, Emanuel Todorov published a paper in Nature Neuroscience [1] that, if you followed the argument carefully, showed something unsettling. He did not dispute the correlations. What he disputed was the inference. Motor cortex sends signals to muscles, and muscles move a limb with particular lengths, masses, and joint configurations. Work through the biomechanics, and almost any reasonable pattern of muscle commands will produce neural activity that looks correlated with movement direction, simply because of the geometry of the arm. Direction tuning might not reflect a neural code for direction. It might be a byproduct of controlling a physical limb.

This should have forced the field to reconsider its foundations. It did not. The debate about what neurons "encode" continued for another two decades. When the field shifted toward dynamical systems in the 2010s, many of the same issues reappeared in new vocabulary, but Todorov's deeper question, about what you can infer from neural correlations given biomechanics, was never cleanly resolved.

Why? Partly incentives: you get published for new results, not for re-examining old ones. But I think there is also an infrastructure problem, and the reason I think this is that I keep running into it.

I develop methods for aligning neural representations across subjects and species [3], which means I regularly collaborate with experimentalists in fields outside my own: motor cortex, songbird vocal production, prefrontal decision-making. Each time I enter a new literature, I go through the same process. I read the big highly cited papers, build a picture of the field, and then slowly discover that some of those landmark results have been substantially weakened by later work with a fraction of the citation count. The critiques are published, peer-reviewed, sometimes technically decisive, but nothing in the literature connects them to the earlier result in a way that would tell you "before you rely on this finding, you should know about these complications." Before I can do my actual research, I have to reconstruct the dependency structure by hand: which results still stand, which have been narrowed, which depend on methodological choices that later turned out to matter more than anyone expected. That reconstruction takes weeks even when I already know roughly what questions to ask. A first-year PhD student entering the same literature would have no reason to suspect the landmark papers were contested at all.

My hypothesis is that this is not mainly a problem of bad incentives. It is a problem of missing infrastructure. Science has become very good at distributing papers, and very bad at storing structured knowledge about the claims those papers contain.

The recycling pattern

Motor cortex is a clean example because the arguments are well documented. Debates restart under new labels: force in the 1960s, direction in the 1980s, kinematics versus kinetics, then representation versus dynamics. Each generation partly absorbed the previous one, but rarely made the inheritance explicit.

Theoretical critiques fail to propagate.*This is not unique to motor cortex. In any field where the dominant framework is easier to use than its replacement, the easier one keeps generating results faster, keeps getting cited, and persists long after the theoretical reasons for preferring it have eroded. Todorov and others showed that once you ask not "what variable do neurons encode?" but "what control law does the circuit implement, given sensory feedback and the biomechanics of the body?" [4], many of the older disputes look much narrower than they seemed. Yet the older framing persists, because it is easier to operationalize. Computing a population vector is straightforward. Building a feedback control model is hard. The easier analysis generates results faster, even if it answers a less well-posed question.

You can see this playing out right now. Churchland and colleagues [5] showed in 2012 that reaching-related neural activity is well described by rotational dynamics, a result that reshaped the field. But then Sauerbrei and colleagues [6] showed cortical activity depends on continuous thalamic input, challenging the idea of autonomous dynamics. Kalidindi and colleagues [7] showed rotational structure is what you'd expect from a feedback controller. Kuzmina and colleagues [8] showed that whether you detect rotations at all depends on your smoothing kernel. And Suresh and colleagues [9] showed the dynamics seen during reaching do not appear during grasping. Each of these should change how you interpret the earlier results. But there is no standard resource that connects them.

Why existing tools are not enough

You might think review articles solve this. They help, but a review captures one author's interpretation at a single moment. Nobody updates it when new results come in. And it narrates the landscape rather than decomposing it: a review might say "the autonomous dynamics view has been challenged," but it won't tell you that the challenge has two logically independent sources, neither of which touches the separate problem of preprocessing sensitivity.

Tools like Semantic Scholar and Scite.ai classify citations, but they work at the level of documents. They can tell you that paper A cites paper B in a contrasting way. They cannot tell you that paper A's claim depends on a 20 ms smoothing kernel, that paper C showed this choice qualitatively changes the result, and that paper D's disagreement is about theoretical interpretation rather than data.

The citation graph compounds the problem. Serra-Garcia and Gneezy [10] found that papers which fail to replicate accumulate more citations than those that replicate. Greenberg [11] traced a single claim through 242 papers and found the vast majority of citations went to supportive studies. A citation graph that cannot distinguish support from challenge is actively misleading when median statistical power is around 21% [12].

What is missing

What would it look like to actually fix this? The key idea is to change the unit of scientific infrastructure from the paper to the claim. Each claim would carry its dependencies explicitly (what data it rests on, what analysis choices it requires, what theoretical commitments it assumes) and have typed connections to other claims: supports, challenges, restricts scope, replicates under different conditions.*A complete system would eventually need additional layers: experimental design rationale, operationalization of variables, interpretation and framing, and consensus tracking. The pilot proposed here starts with claims and dependencies, where the gap is most acute.

Consider the rotational dynamics debate. Churchland et al. 2012 would not appear as a single entry but be decomposed into constituent claims, each carrying its dependencies. Kuzmina's preprocessing result would attach to the specific dependency on the smoothing kernel. That flag would propagate to every downstream study whose inference depends on similar preprocessing. Studies that reached related conclusions through optogenetic perturbation or feedback-control modeling would be unaffected, because their evidence goes through different dependencies.

This selective propagation is the core of the idea. Right now the field has two modes: nothing happens when a critique is published, or some human reader holds the entire web of dependencies in their head. A claim-dependency graph would create a third option, where challenges move along actual epistemic connections rather than relying on social memory.

But the implications go further. If five labs have tested a claim using different preprocessing, species, and task designs, that convergence is far stronger evidence than any single paper, but no existing system makes it visible. Follow-up work that tightens the scope of a claim currently has almost no career value because it is not "new"; in a claim layer, it strengthens an edge, and that contribution is visible. Null results gain a role: a study that tests a claim under new conditions and finds it does not hold is not a failure but a scope restriction, and the graph would represent that as a real object. And claims that cannot both be true, because they assume incompatible preprocessing or theoretical commitments, would be flagged by the structure of the graph itself.

This is what I mean by a "lean for science." Not a proof assistant. Empirical science is too messy for that. But a system where claims must be situated relative to other claims, and where contradictions become visible by default rather than buried by convention.

Realizing any of this requires solving a genuinely hard design problem: what counts as a "claim"? Churchland et al. 2012 contains at least three separable claims, and reasonable people could decompose it differently. Too coarse and you lose the dependency structure; too fine and experts cannot verify entries in reasonable time. The schema has to be tested empirically.

The approach I'd propose has three layers. Language models extract candidate claims, dependencies, and edges from paper text. A graph-level algorithm propagates sensitivity flags along dependency chains, detects inconsistencies between claims with incompatible assumptions, and surfaces clusters of claims that converge through independent paths. Domain experts adjudicate the outputs of both layers. The schema would cover a fixed set of dependency types (data source, species, task, preprocessing, statistical test, model class, theoretical commitment) and edge types (supports, depends-on, challenges, restricts-scope, replicates). Whether it is expressive enough is one of the things the pilot would test.

The experiment

Why now? Standardized neural data archives have reached critical mass: DANDI hosts over 1,000 datasets, and the Neural Latents Benchmark provides preprocessed recordings from the motor cortex experiments at issue.*Before language models, structured extraction from papers meant either building narrow rule-based parsers or paying domain experts to annotate each paper from scratch. Language models can propose a candidate decomposition that is wrong often enough to need expert review, but right often enough that the expert's job shifts from generation to correction. That shift is what makes a 400-paper corpus tractable. Language models have made structured extraction from scientific text much cheaper, making a few-hundred-paper corpus tractable for a small team. And AI-assisted writing is accelerating paper production without adding structure [13], making the problem worse faster than anyone is building tools to address it.

I propose building a claim-dependency graph for a relatively small, well-documented field (motor neuroscience is the example I've developed here, but the approach is not specific to it). Assemble a corpus of roughly 300 to 500 core papers. Define the schema. Use language models to propose candidate entries; have domain experts adjudicate. Release the graph publicly.

The most informative test is a backtest. Build the graph using only papers published before some cutoff, say 2018. Then ask: does the dependency structure flag the problems the field took years to notice? Does it surface the preprocessing sensitivity that Kuzmina published in 2024, the tension between autonomous and input-driven dynamics before Sauerbrei 2020, the possibility that reaching results might not generalize to grasping? If yes, the infrastructure would have saved years of confused debate. If no, either the schema is too coarse or the dependencies that matter were not visible in the earlier literature, and the bottleneck is elsewhere.

No existing institution is built to maintain this. Universities do not reward infrastructure work. Grant panels evaluate novelty. Publishers have no reason to flag contradictions in their own product. It would require a dedicated team, probably something like a Focused Research Organization.*The GenBank model is instructive. GenBank succeeded because ACGT strings are canonical, journal mandates required deposition, and government funding sustained it indefinitely. Scientific claims are harder: ambiguous, contested, expressed in natural language. The LLM-plus-expert-plus-algorithm hybrid is the first approach that might bridge this gap at reasonable cost.

I want to be honest about how this could fail. The schema may be too rigid to capture the distinctions that matter. The graph may capture real structure but fail the backtest, surfacing only tensions already obvious from the pre-2018 literature. Or the backtest may succeed for known cases but fail to generalize, because we built it knowing what to look for. Any of these would be worth knowing.

Motor neuroscience is a good pilot domain because it has a bounded literature, standardized data formats, and active disputes with identifiable dependency structure. But it is not unique. Gene Ontology gave biology a shared vocabulary. The Protein Data Bank gave structural biology a shared archive. Lean gave mathematics a shared verification layer. Neuroscience has built the data archives and the computation standards. What it still lacks is the glue: a structured, queryable layer connecting claims to evidence to data to code. Building that layer for one subfield is a tractable experiment. If it works, it becomes a template.

References

1. E. Todorov, "Direct cortical control of muscle activation in voluntary arm movements: a model," Nature Neuroscience, vol. 3, pp. 391–398, 2000.
2. A. P. Georgopoulos, A. B. Schwartz, R. E. Kettner, "Neuronal population coding of movement direction," Science, vol. 233, pp. 1416–1419, 1986.
3. A. Ramot, F. H. Taschbach, Y. C. Yang, et al., "Motor learning refines thalamic influence on motor cortex," Nature, 2025.
4. S. H. Scott, "Optimal feedback control and the neural basis of volitional motor control," Nature Reviews Neuroscience, vol. 5, pp. 532–546, 2004.
5. M. M. Churchland, J. P. Cunningham, et al., "Neural population dynamics during reaching," Nature, vol. 487, pp. 51–56, 2012.
6. B. A. Sauerbrei, J.-Z. Guo, et al., "Cortical pattern generation during dexterous movement is input-driven," Nature, vol. 577, pp. 386–391, 2020.
7. H. T. Kalidindi et al., "Rotational dynamics in motor cortex are consistent with a feedback controller," eLife, vol. 10, e67256, 2021.
8. E. Kuzmina, D. Kriukov, M. Lebedev, "Neuronal travelling waves explain rotational dynamics in experimental datasets and modelling," Scientific Reports, vol. 14, 3566, 2024.
9. A. K. Suresh, J. M. Goodman, et al., "Neural population dynamics in motor cortex are different for reach and grasp," eLife, vol. 9, e58848, 2020.
10. M. Serra-Garcia, U. Gneezy, "Nonreplicable publications are cited more than replicable ones," Science Advances, vol. 7, eabd1705, 2021.
11. S. A. Greenberg, "How citation distortions create unfounded authority: analysis of a citation network," BMJ, vol. 339, b2680, 2009.
12. K. S. Button et al., "Power failure: why small sample size undermines the reliability of neuroscience," Nature Reviews Neuroscience, vol. 14, pp. 365–376, 2013.
13. A. Narayanan, S. Kapoor, AI Snake Oil: What Artificial Intelligence Can Do, What It Can't, and How to Tell the Difference, Princeton University Press, 2024.