In arXiv:2410.03595v1 (preprint under review; PDF stamp 4 Oct 2024), the neuron-to-neuron computation is a written characterization of the Sherringtonian view, not a measurement on a language-model checkpoint. Appendix A states that, at an algorithmic level, cognition is a network of nodes (neurons) with weighted connections (synapses): each neuron receives signals from other neurons, processes those inputs through a neural transfer function, and transmits the resulting signal to connected neurons. The v1 text does not print a symbolic form, weights, threshold, or activation for that function, and the authors do not use this description as their Chain-of-Thought method.
Key Claims
- Written exposition, not a result: Appendix A of v1 describes the computation in prose and attributes the view to Sherrington (1906), Barlow (1953), and circuit-level explanation as in Mogenson (2018).
- Sequence as written: inputs from other neurons, a distinct computational transformation called a neural transfer function, then transmission of the resulting signal across weighted connections. Cognition, on this view, is those per-neuron computations plus signal flow.
- Not observed in this paper: no model, layer, intervention, or metric is attached to that sentence. It is not a causal or correlational finding about transformer units.
- Author stance, separately labeled: the preprint says it adopts a Hopfieldian, representation-space account instead, and it classifies some other neuron-level LLM studies as Sherringtonian. That classification is interpretation, not a fitted transfer function.
In This Guide
Appendix A’s algorithmic account
arXiv:2410.03595v1, “Understanding Reasoning in Chain-of-Thought from the Hopfieldian View” (Hu, Liu, Yang, Chen, Tan, Ali, Li, and Wang), points to this description from a Section 3 footnote: “See Appendix A for an introduction to the Sherringtonian view.” The transfer-function wording itself is in that appendix, under “The Sherringtonian View,” not in the main-text contrast and not in the paper’s later Representation-of-Thought updates.
The appendix first states what the view is supposed to explain. Unlike the Hopfieldian perspective, “the Sherringtonian view (Sherrington, 1906; Barlow, 1953) of cognitive explanation emphasizes the importance of direct neuron-to-neuron connections in the brain.” It “posits that the primary explanation for cognition lies in the specific interactions between neurons and the computations these neurons perform within well-defined circuits (Mogenson, 2018).”
At an algorithmic level, the Sherringtonian view conceptualizes cognition as networks of nodes (neurons) with weighted connections (synapses) between them. In this framework, neurons perform distinct computational transformations on the signals they receive from other neurons in the network. Cognitive processes are described by how individual neurons receive inputs, process these inputs through neural transfer functions, and transmit the resulting signals to connected neurons. Thus, cognition is explained through the computations occurring at the level of individual neurons and the signal flow across their connections.
That paragraph is the paper’s full written neuron-to-neuron computation. Read in order, it specifies four objects and one explanatory claim:
- Nodes are neurons. Weighted connections are synapses. The paper equates those pairs parenthetically; it does not add a separate weight-learning rule.
- Inputs are signals received from other neurons in the network, not an unspecified external feature vector and not the Chain-of-Thought “stimuli” defined later for the authors’ own method.
- The local computation is a “distinct computational transformation,” which the next sentence names as processing inputs “through neural transfer functions.” The two phrases describe that same step. “Distinct” is the paper’s adjective; v1 does not say whether every neuron must have a unique function.
- The output of that function is transmitted to connected neurons. Cognition, on this view, is the set of those individual-neuron computations together with signal flow across the connections.
As used in this preprint, the Sherringtonian view is an algorithmic account: explanation is located in neuron-to-neuron connections inside well-defined circuits, with each neuron applying a transfer function to incoming signals. Barlow enters only as a citation for that view, Horace B. Barlow, “Summation and inhibition in the frog’s retina” (1953), listed beside Sherrington’s 1906 scratch-reflex paper. v1 does not quote either source on transfer functions, so the phrase “neural transfer functions” should be treated as this preprint’s wording.
What the passage does not define
The appendix states the computation in prose. In the v1 text used here, that subsection does not define a symbol for the transfer function, a weighted sum, a threshold, an activation, a time index, or an update schedule. It does not say whether the map is linear, whether weights are signed, or whether transmission is feedforward or recurrent. Filling any of those in would go beyond the written account.
Equations that do appear on the same appendix page are not this computation. After the Sherringtonian subsection, the preprint defines zero-shot and few-shot Chain-of-Thought generation as an argmax over response probability given a query, a prompt or demonstration set, and model weights. Those are prompting likelihoods. They do not specify a synaptic transfer function, and they should not be read as the missing formula for the neuron-level step.
No checkpoint, layer index, or intervention is identified for this description, because the passage is not an empirical claim. Do not substitute the paper’s later Hopfieldian operations, activation differences, a principal-component direction, a score against that direction, or an additive edit of hidden states, for the neural transfer function in Appendix A. Those operations are representation-space procedures in the authors’ own method, not the synapse-weighted node computation quoted above.
Text extraction of the PDF does not replace a figure-level check. The Sherringtonian subsection, as retained, is continuous prose and does not contain a displayed transfer-function equation. This article does not claim that an unextracted graphic elsewhere in a later version is absent; it claims only that v1’s written introduction of the view does not supply one.
How the authors separate this view from their method
The preprint’s project is Chain-of-Thought interpretation under a Hopfieldian framing. On the first pages it says the Sherringtonian view focuses on specific connections between neurons, whereas the Hopfieldian view emphasizes distributed computation across neural populations, with information not encoded by a single neuron. It calls the Hopfieldian view “generally considered more advanced,” especially for distributed computation and network dynamics, citing Barack and Krakauer (2021). A Section 3 footnote also points to Bechtel (2007) for a detailed comparison. That ranking is the authors’ citation-backed interpretation, not a result of an intervention in this preprint.
Their own mapping therefore does not implement Appendix A. They align Chain-of-Thought prompts with stimuli, model activations with neural populations, and low-dimensional subspaces with representation spaces, and they describe reasoning as movement between those spaces. They explicitly set aside individual cells and circuits in the Hopfieldian picture they adopt. Leaving attractors and state space as future work is a limit on that Hopfieldian program; it is not a specification of the Sherringtonian transfer function.
Two uses of “circuit” in the paper should stay separate. Appendix A’s circuits are the “well-defined circuits” of the cited neuroscience framing (Mogenson, 2018): networks of neurons and weighted synapses. In related work, the authors also mention mechanistic interpretability that explains networks by circuits and node-to-node connections, citing Olah et al. (2020) among others, and they write that those methods cannot be directly used to interpret Chain-of-Thought reasoning. That literature mention is not a measurement of Appendix A’s transfer functions.
The authors do apply the Sherringtonian label to other LLM studies: Dutta et al. (2024) on neural sub-structures for Chain-of-Thought in Llama-2-7B, and Rai and Yao (2024) on feed-forward neurons and arithmetic reasoning in Llama-2-7B. They write that both are grounded in the Sherringtonian view of neural activity, and that their own work instead focuses on representations rather than individual neurons, including analyses they say can extend to larger models such as Llama-2-70B. That is author classification of other papers. v1 does not show that those studies use Appendix A’s transfer-function wording, and this article does not import their dimensions or numerical results.
The evidence for the computation above is the v1 PDF text, principally Appendix A, cross-checked against the Section 3 footnote that sends the reader there and against the opening contrast that does not yet use the phrase “neural transfer functions.” Production of that evidence is exposition plus citation, not an experiment: there is no dataset, prompt condition, layer set, control, ablation, or accuracy metric for the transfer-function sentence. It establishes how this preprint describes the Sherringtonian computation. It does not establish that biological neurons, or units in any named model, implement a particular function, and it does not re-derive Sherrington (1906), Barlow (1953), or Mogenson (2018).
Limits and what remains unknown
Only arXiv:2410.03595v1, marked “Preprint. Under review,” was used. Whether a later version revises Appendix A is unknown. The historical sources were not re-read for this account, so it is unknown whether Sherrington, Barlow, or Mogenson use “neural transfer function” in the sense printed here. The reference-list titles identify those citations as physiology and textbook sources, not as language-model evaluations.
Nothing in the passage is a correlation, a causal intervention, or a replication. The paper reports no test that would confirm or falsify the transfer-function description, and no such test is proposed here. Readers who need a fitted neuron-to-neuron function, a weight matrix, or a circuit in a specific checkpoint will not find it in this written account; treating the authors’ Hopfieldian edits as that function would change the claim.
- Appendix A’s algorithmic account
- What the passage does not define
- How the authors separate this view from their method
- Limits and what remains unknown
Sources
