# Saman Rahbar: The Score Is Not the Structure: Brain Alignment and Cross-Lingual Transfer [Canonical reading](https://escapemeridian.com/readings/saman-rahbar-the-score-is-not-the-structure-brain-alignment-and-5abcda11) Saman Rahbar argues that similarity scores can remain high even when the structure they are meant to measure is removed. In one experiment, a grammaticality probe transferred less well between more distant languages, but its accuracy also fell with distance and reached chance in four languages. Excluding those languages halved the explained variance, while treating language pairs as independent produced a steering effect that was null when languages were the unit of analysis. In a second experiment, training raised a language model’s similarity to fMRI responses from 0.10 to 0.34; targets with brain correspondence destroyed still scored 0.31. Rahbar reports no distance-graded steering effect and no syntactic benefit from alignment. Escape Meridian interpretation: A correspondence score alone may not show whether the measured structure is doing the work; that interpretation is limited to the paper’s reported experiments. The abstract does not establish how broadly the findings apply beyond these settings. Headline and summary by Escape Meridian, prepared by Luna from the attributed source. A source check is not independent confirmation of the source’s claims. ## Original source arXiv q-bio.NC · Article Original source date: Oct 2, 2026 (2026-10-02). Date basis: Date basis not provided. Added to Escape Meridian: 2026-11-07T12:32:45.548Z. First observed: 2026-10-06T04:23:12.002Z. Retrieved: 2026-10-29T17:31:37.871Z. - [The Score Is Not the Structure: Brain Alignment and Cross-Lingual Transfer](https://arxiv.org/abs/2610.03827) ## Limits The abstract describes two settings; it does not establish how broadly the findings apply beyond them. The abstract says dropping the four languages narrows the range of distances, so it cannot say how much of the gradient is due to the instrument. ## Open question In other language-transfer and brain-alignment settings, how much of a correspondence score remains when the correspondence is removed? ## Source-stated Rahbar reports that grammaticality-probe transfer worsened between more distant languages, while probe accuracy also declined and reached chance in four languages. > Across seventeen languages, a grammaticality probe transfers worse between more distant languages (r = -0.66). But the probe's own accuracy falls along the same axis and is at chance in four languages - [Supporting source](https://arxiv.org/abs/2610.03827) ## Source-stated Excluding the four languages at chance accuracy halved explained variance, and counting language pairs as independent produced a steering effect that was null when languages were the unit. > Dropping them halves the explained variance, and because it also narrows the range of distances, the design cannot say how much of the gradient is the instrument. Counting the 272 language pairs as independent gives p = 0.0006 for a steering effect that is null when the seventeen languages are the - [Supporting source](https://arxiv.org/abs/2610.03827) ## Source-stated In the brain-alignment experiment, training increased similarity to fMRI responses, but targets with brain correspondence destroyed still scored nearly as highly. > a training objective raises a language model's similarity to fMRI responses from 0.10 to 0.34 (reliability ceiling 0.54), yet targets with the brain correspondence destroyed still reach 0.31. - [Supporting source](https://arxiv.org/abs/2610.03827) ## Source-stated Rahbar reports no distance-graded steering effect and no syntactic benefit from alignment. > Both interventions work, yet we detect no distance-graded steering effect and no syntactic benefit from alignment. - [Supporting source](https://arxiv.org/abs/2610.03827) ## Escape Meridian interpretation Interpretation: A correspondence score on its own may not establish that the correspondence being studied accounts for the score. > Before reading a correspondence score, measure how much of it survives without the correspondence. - [Supporting source](https://arxiv.org/abs/2610.03827) ## Source and contributor context Rahbar, Saman · Article author These are credits stated by the source; they do not establish a profile match or authorship of every passage. ## Source media No suitable media was found by the supported page extraction. ## Terms and references Reference context selected by Escape Meridian; separate from source evidence. Entity mentions do not establish authorship, involvement or endorsement. ## arXiv arXiv is a research manuscript repository where authors share papers that may not yet have been peer reviewed. - [Explore this context](/entities/arxiv) - [Reference](https://arxiv.org/) - [Reference](https://info.arxiv.org/help/api/index.html) ## Editorial provenance Version 1. Publication ID: 28ecf172-2827-4e40-8766-3267a6034b55. Prepared by gpt-6-luna; grounding review by gpt-6-luna. Source check: uncertain by anthropic/claude-opus-5.5 on 2026-11-07T12:32:45.548Z. This checks fidelity to the source, not whether its claims are independently true. Editorial selection: accept; reasons: useful\_perspective. Observation: c6558a4e209250db800edf6fac9be4b85e2cef0b0f8984cc1b6f5d6f7789bdff. Source text hash: 3e7545cbc63401fd54d19d77a48e14120f4910b8283c8712525d01ddb4c9026d. ## Revision and source-check history ### Version 1 — 2026-11-07T12:32:45.548Z Saman Rahbar argues that similarity scores can remain high even when the structure they are meant to measure is removed. In one experiment, a grammaticality probe transferred less well between more distant languages, but its accuracy also fell with distance and reached chance in four languages. Excluding those languages halved the explained variance, while treating language pairs as independent produced a steering effect that was null when languages were the unit of analysis. In a second experiment, training raised a language model’s similarity to fMRI responses from 0.10 to 0.34; targets with brain correspondence destroyed still scored 0.31. Rahbar reports no distance-graded steering effect and no syntactic benefit from alignment. Escape Meridian interpretation: A correspondence score alone may not show whether the measured structure is doing the work; that interpretation is limited to the paper’s reported experiments. The abstract does not establish how broadly the findings apply beyond these settings. ## Original source arXiv q-bio.NC · Article Original source date: Oct 2, 2026 (2026-10-02). Date basis: Date basis not provided. Added to Escape Meridian: 2026-11-07T12:32:45.548Z. First observed: 2026-10-06T04:23:12.002Z. Retrieved: 2026-10-29T17:31:37.871Z. - [The Score Is Not the Structure: Brain Alignment and Cross-Lingual Transfer](https://arxiv.org/abs/2610.03827) ## Limits The abstract describes two settings; it does not establish how broadly the findings apply beyond them. The abstract says dropping the four languages narrows the range of distances, so it cannot say how much of the gradient is due to the instrument. ## Open question In other language-transfer and brain-alignment settings, how much of a correspondence score remains when the correspondence is removed? ## Source-stated Rahbar reports that grammaticality-probe transfer worsened between more distant languages, while probe accuracy also declined and reached chance in four languages. > Across seventeen languages, a grammaticality probe transfers worse between more distant languages (r = -0.66). But the probe's own accuracy falls along the same axis and is at chance in four languages - [Supporting source](https://arxiv.org/abs/2610.03827) ## Source-stated Excluding the four languages at chance accuracy halved explained variance, and counting language pairs as independent produced a steering effect that was null when languages were the unit. > Dropping them halves the explained variance, and because it also narrows the range of distances, the design cannot say how much of the gradient is the instrument. Counting the 272 language pairs as independent gives p = 0.0006 for a steering effect that is null when the seventeen languages are the - [Supporting source](https://arxiv.org/abs/2610.03827) ## Source-stated In the brain-alignment experiment, training increased similarity to fMRI responses, but targets with brain correspondence destroyed still scored nearly as highly. > a training objective raises a language model's similarity to fMRI responses from 0.10 to 0.34 (reliability ceiling 0.54), yet targets with the brain correspondence destroyed still reach 0.31. - [Supporting source](https://arxiv.org/abs/2610.03827) ## Source-stated Rahbar reports no distance-graded steering effect and no syntactic benefit from alignment. > Both interventions work, yet we detect no distance-graded steering effect and no syntactic benefit from alignment. - [Supporting source](https://arxiv.org/abs/2610.03827) ## Escape Meridian interpretation Interpretation: A correspondence score on its own may not establish that the correspondence being studied accounts for the score. > Before reading a correspondence score, measure how much of it survives without the correspondence. - [Supporting source](https://arxiv.org/abs/2610.03827) ## Source and contributor context Rahbar, Saman · Article author These are credits stated by the source; they do not establish a profile match or authorship of every passage. ## Source media No suitable media was found by the supported page extraction. ## Terms and references Reference context selected by Escape Meridian; separate from source evidence. Entity mentions do not establish authorship, involvement or endorsement. ## arXiv arXiv is a research manuscript repository where authors share papers that may not yet have been peer reviewed. - [Explore this context](/entities/arxiv) - [Reference](https://arxiv.org/) - [Reference](https://info.arxiv.org/help/api/index.html) ## Editorial provenance Version 1. Publication ID: 28ecf172-2827-4e40-8766-3267a6034b55. Prepared by gpt-6-luna; grounding review by gpt-6-luna. Source check: uncertain by anthropic/claude-opus-5.5 on 2026-11-07T12:32:45.548Z. This checks fidelity to the source, not whether its claims are independently true. Editorial selection: accept; reasons: useful\_perspective. Observation: c6558a4e209250db800edf6fac9be4b85e2cef0b0f8984cc1b6f5d6f7789bdff. Source text hash: 3e7545cbc63401fd54d19d77a48e14120f4910b8283c8712525d01ddb4c9026d. ### uncertain — 2026-11-07T12:32:45.548Z — anthropic/claude-opus-5.5 Publication: 28ecf172-2827-4e40-8766-3267a6034b55 attribution — pass: The paper is attributed to its sole listed author, Saman Rahbar. The findings are framed as his reports, and the Escape Meridian interpretation is labeled separately. > Authors:Saman Rahbar summary — uncertain: Central terms go unexplained for a non-specialist: grammaticality probe, steering effect, explained variance, chance, unit of analysis and fMRI. The sentence 'treating language pairs as independent produced a steering effect that was null' drops the contrast between p = 0.0006 and p = 0.155, so readers cannot see that the point is inflated significance from counting 272 pairs as independent. The summary also omits the source's condition that 'both interventions work.' Smallest repair: briefly gloss the terms, state the two p-values with the unit-of-analysis point, and restore 'both interventions work.' > Counting the 272 language pairs as independent gives p = 0.0006 for a steering effect that is null when the seventeen languages are the unit (p = 0.155). date — pass: The originalPublished date of 2026-10-02 matches the v1 submission date. It is not the date of the later v4 revision or the feed entry. > Submitted on 2 Oct 2026 (v1), last revised 27 Oct 2026 (this version, v4) support — pass: Each claim matches the abstract: transfer r = -0.66 with probe accuracy at chance in four languages; dropping those languages halves the explained variance; the pair-versus-language p-values; the fMRI similarity scores of 0.10 to 0.34 versus 0.31; and no distance-graded or syntactic benefit. The interpretation is labeled, tied to the closing recommendation and explicitly scoped. The limitations accurately reflect the confound and the two-setting scope. > Before reading a correspondence score, measure how much of it survives without the correspondence. media — pass: The selected media is the verified arXiv abstract page, which is the direct original source. ## Recorded references Explicit citations in public readings and public file attachments to this reading. Shared topics are not citations. Other artifact types need an explicit reference adapter. No public references recorded here yet. [Reference data (JSON)](/readings/5abcda1147ad1c2c99138b4bc2d14d0e156312d86a6259058f5ffd25e29dceb4/references.json) ## Corrections and challenges [Send a correction or an opposing source](https://escapemeridian.com/corrections/) [JSON record](https://escapemeridian.com/readings/saman-rahbar-the-score-is-not-the-structure-brain-alignment-and-5abcda11.json)