No country for old linguists: LLM-brain alignment underdetermines neural computation
cs.CL
Quick verdict
Commentary: LLM-brain alignment constrains but cannot identify neural mechanisms; sharp critique, but proposed algebraic alternative is asserted, not tested.
Summary
This is a conceptual commentary responding to Nastase et al. (2026), who argue that LLM-brain alignment supports viewing large language models as mechanistic models of natural language. Murphy accepts that alignment research is valuable but argues that its inferential reach is narrower than claimed. Section 2 formalizes the concern: in an encoding model Y = WX + ε, an unconstrained linear map W preserves only collinearity and parallelism, not the distances and angles that 'shared geometry' normally means; a common-input problem (both LLM and brain respond to the same stimulus) can generate alignment without shared computation; and, following Guest and Martin (2023), inference from prediction alone amounts to affirming the consequent. Section 3 draws an analogy with surprisal (Slaats and Martin, 2025): a strong predictor need not identify latent structure. Section 4 argues that the mechanistic evidence cited by Nastase et al. (Kumar et al., 2024; Zada et al., 2024) characterizes artificial circuitry, not cortical operations. Section 5 contends that 'high-dimensional population code' is a format rather than a theory, faces a falsifiability problem, and confuses decodability with computationally addressable structure; the author proposes algebraic structure-preservation conditions and references his own ROSE model. Main limitations: no new data or simulations, the target paper's claims are engaged only through quotations, and the proposed alternative is asserted rather than tested.
Review complete
Assessment at a glance
5
risk flags identified
● high 0
● medium 4
● low 1
5 strengths · 5 weaknesses · 5 author questions
AI text signal = Binoculars zero-shot detector score over 16 sampled passages (Fast-DetectGPT co-reported) — absence does not prove human authorship, and a flag does not imply misconduct.
Levels reflect verification complexity — not a quality score. Evolve never rates your paper.
Key claims under review
- Representational alignment between LLMs and brains can constrain mechanistic hypotheses but does not by itself identify a mechanism (Introduction; Section 2).
- Regression-based encoding evidence with an unconstrained linear map establishes only a predictive linear relation, not shared representational geometry, because arbitrary linear maps preserve collinearity and parallelism but not distances or angles (Section 2).
- Apparent LLM-brain alignment may be inherited from shared stimulus structure rather than shared computation (common-input problem, Section 2).
- 'High-dimensional population code' is a representational format, not a computational theory, and decodability by an external analyst does not imply the structure is computationally addressable by the brain itself (Section 5).
- Formal algebraic structure-preservation conditions (e.g., non-associative composition) can rule out whole classes of candidate neural implementations, as pursued in the author's ROSE program (Section 5).
Strengths
Formally precise core objection
Section 2 states the invariance point exactly: an unrestricted W in Y = WX + ε guarantees survival only of relations invariant under arbitrary linear maps — collinearity and parallelism (∥Wx_i − Wx_j∥ ≠ ∥x_i − x_j∥ in general) — correctly separating 'predictive linear relation' from 'shared geometry'.
Common-input confound is a live methodological issue
Section 2's S → X_LLM, S → Y_brain point, backed by Schönmann et al. (2026, eLife), identifies a genuine confound in naturalistic encoding designs of the kind used by Kumar et al. (2024) and Zada et al. (2024).
Decodability vs. computational addressability distinction
Constructive criterion, not pure negation
Falsifiability pressure on the field
Weaknesses
No new evidence by design
Self-referential support loop
Critique partially overgeneralized
Contested linguistic premise treated as given
Rhetorical framing
Full Review Report
$1.99 · one-timeUnlock the complete preliminary analysis Evolve produced for this paper.
Report preview
Preliminary Review: 'No country for old linguists: LLM–brain alignment underdetermines neural computation' (Murphy, arXiv 2609.03160, cs.CL)
The remaining sections cover Methodology Assessment, Statistical Evidence, Results & Reproducibility, Limitations & Risks, and Verdict & Recommendations.
- All 5 strengths & 5 weaknesses, analyzed in detail
- 5 risk flags with severity, section & reviewer questions
- Methods & statistical evidence assessment
- 5 questions a peer reviewer would ask
- Full six-dimension evaluation report & recommendations
One-time purchase · this paper only
One-time purchase, this paper only. AI-generated preliminary analysis — not peer review.
Frequently asked questions
What is the main argument of Murphy's 'No country for old linguists' paper?
It argues that LLM-brain alignment studies (specifically Nastase et al., 2026) show that a model predicts brain activity, but this does not by itself establish shared geometry, shared computational principles, or a neural mechanism. Prediction constrains explanation but is not itself explanation.
Is this paper an empirical study with new data or experiments?
No. It is a conceptual commentary with no datasets, analyses, or code — the author states no datasets were generated or analyzed. Its arguments are logical and draw on prior studies such as Kumar et al. (2024), Zada et al. (2024), and Slaats and Martin (2025).
What does the author propose instead of LLM-brain alignment?
He advocates formal, algebraic structure-preservation conditions: a neural implementation is admissible only if its dynamics preserve the relations that define a linguistic capacity (e.g., non-associative constituent grouping). He develops this approach in his 'ROSE' model of a universal neural grammar (Murphy, 2025).
Why is the paper titled 'No country for old linguists'?
The title wryly references the common claim that formal, symbolic linguistics is obsolete in the LLM era. The paper argues the opposite: formal linguistic structure remains essential for constraining which neural mechanisms could implement language, and population-code descriptions alone do not supply that constraint.
About this review
This is an AI-generated preliminary review by Evolve.Pub, produced from the openly available text of the paper. It is not peer review and does not represent the authors' views. The canonical record of the paper is at the publisher link above.
- Equation subscripts are mangled by PDF extraction (e.g., 'Y!"#$% = WX&&' + ε'); all formulas were reconstructed from context (Y = WX + ε; ‖Wx_i − Wx_j‖ ≠ ‖x_i − x_j‖; surprisal I(w_t|c_t); N(M(x,y)) ≃ G(N(x), N(y))).
- Figure 1's graphical content is an image not present in the extracted text; only its caption was reviewed.
- The target paper (Nastase et al., 2026) was not provided; its position could be assessed only through Murphy's quotations, which were not verifiable against the original.
Go deeper on this paper
Wrote a paper? Stress-test it before reviewers do.
Upload your manuscript for the same structured scrutiny — reasoning, statistics, evidence, figures, and reproducibility.