Pith. sign in

REVIEW 4 major objections 2 minor 40 references

Bridging Brains and Models: MoE-Based Functional Lesions for Simulating and Rehabilitating Aphasia

T0 review · 4 major / 2 minor · reviewed 2026-08-06 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read This paper claims that disabling syntax experts in a mixture-of-experts language model reproduces Broca's aphasia, disabling semantics experts reproduces Wernicke's aphasia, and retraining the remaining experts restores function.

desk verdict Interesting MoE-lesion idea, but the abstract alone doesn't rule out circular expert selection—needs the full text. read the letter →

arxiv 2508.04749 v1 pith:N7BI6TPA submitted 2025-08-06 q-bio.NC

Yifan Wang , Jingyuan Sun , Jichen Zheng , Yunhao Zhang , Chunyu Ye , Jixing Li , Chengqing Zong , Shaonan Wang This is my paper
classification q-bio.NC
keywords aphasiamixture-of-expertslargelanguagemodelsfunctionallesionBroca'sWernicke'scomputationalrehabilitationdisorders
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

This paper tries to establish that a modular large language model can serve as a computational testbed for brain injury. Selectively disabling the experts responsible for syntax produces language behavior resembling Broca's aphasia, while disabling the experts responsible for semantics produces behavior resembling Wernicke's aphasia, and the lesioned outputs are validated against real patient speech. It then shows that freezing the damaged experts and retraining only the healthy experts on conversational data restores a substantial amount of linguistic function. If true, this gives researchers an in silico analogue of both aphasia and rehabilitation, with the potential to explore therapy strategies computationally.

What carries the argument

The central object is the Mixture-of-Experts (MoE) layer, a neural network component with several smaller sub-networks, or experts, that specialize in different inputs and are selected by a routing mechanism. The method maps neural injury onto this architecture: a functional lesion disables a selected expert or group of experts, and rehabilitation is simulated by freezing those damaged experts while retraining the intact ones on conversational data. The argument turns on the assumption that the experts selected as syntax and semantics experts are genuinely specialized, so that disabling them produces selective, not general, language deficits.

What would settle it

A direct test would rerun the lesioning pipeline with experts chosen at random instead of by syntax/semantics specialization; if random ablations produce equally strong aphasia-like profiles, the claimed specificity fails. A second check would measure whether a syntax-expert lesion impairs semantic probes as much as grammatical ones and vice versa: true dissociation requires selective, not uniform, damage. The rehabilitation claim could also be falsified by showing that retraining with no frozen experts, or on non-conversational text, restores just as much function.

Watch

Extended reading notes

Core claim

The core claim is a functional double dissociation in a Mixture-of-Experts (MoE) language model: certain experts are specialized for syntax, others for semantics, and ablating the syntax group yields effortful, grammatically broken speech similar to Broca's (non-fluent) aphasia, whereas ablating the semantics group yields fluent but semantically empty speech similar to Wernicke's (fluent) aphasia. The paper further claims that the impairment is not permanent: with the damaged experts frozen and the intact experts retrained on conversational data, the model regains significant linguistic function. The authors position this as a computational analogue of rehabilitation and as a clinically rele

Load-bearing premise

The load-bearing premise is that the experts labeled syntax and semantics are genuinely and cleanly specialized, because if disabling them only degrades language broadly, the match to Broca's and Wernicke's aphasia is an artifact of how the experts were selected rather than evidence about aphasia.

Editorial extensions

If this is right

  • Structured ablations in an MoE language model can generate candidate aphasic speech patterns that can be compared directly with patient transcripts, providing a data-rich in silico model of aphasia subtypes.
  • The rehabilitation result implies that recovery is carried by intact experts re-learning to compensate, rather than by repair of the damaged components, a concrete prediction about post-injury plasticity.
  • The framework makes it possible to screen retraining conditions (data type, amount, frozen vs. unfrozen modules) computationally before deciding which rehabilitation protocols to test clinically.
  • The syntax/semantics dissociation suggests that expert modularity in LLMs is functionally meaningful and not merely an implementation detail.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • The paper does not state how the syntax and semantics specialization of the experts was established; if the same behavioral probes used for validation also selected the experts, part of the dissociation could be circular. A cleaner version would choose experts by an independent criterion (e.g., routing patterns) and validate on held-out probes.
  • A natural extension left implicit is to vary the number and combination of lesioned experts, predicting that partial or mixed lesions should produce graded deficits and possibly patterns closer to milder or mixed aphasias.
  • The rehabilitation result would be sharper with control conditions—lesioning randomly chosen experts, or retraining without freezing—to show that the recovery is specific to functional reorganization rather than a generic response to more conversational text.
  • If the approach extends, other language disorders with a clearer computational signature (e.g., conduction aphasia or anomia) could be modeled by choosing the corresponding expert subsets, provided the model's specialization structure is known first.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

4 major / 2 minor

Summary. The paper proposes a computational model of aphasia using a Mixture-of-Experts (MoE) language model. It claims that selectively disabling experts that are 'functionally-specialized for syntax or semantics' produces distinct impairments resembling Broca's and Wernicke's aphasia, that these outputs are validated against real patient speech, and that freezing the damaged experts and retraining the intact ones on conversational data 'restores significant linguistic function.' These claims are presented in the abstract. The supplied full text is corrupted (mojibake), so no methods, results, or analyses are readable; the evaluation below is necessarily based on the abstract and the internal consistency of the claims.

Significance. If the central claims hold, this work would introduce a novel in silico platform for studying aphasia subtypes and rehabilitation, potentially bridging computational linguistics and clinical research. The idea of using MoE expert ablation as a structured lesion model is interesting and testable. However, the abstract alone does not provide sufficient evidence to establish the claims. The most important scientific question is whether the expert specialization is truly functional and independent of the validation probes; without this, the dissociation may be partly definitional. The manuscript also lacks quantitative comparisons to patient data and any statistical support for the rehabilitation effect. These issues are fixable, making the central hypothesis defensible in principle, but the current version cannot be accepted.

major comments (4)
  1. [Abstract] The central premise is that MoE experts are 'functionally-specialized' for syntax or semantics, but the abstract does not state how this specialization was determined. If the experts were selected using the same syntactic/semantic probes later used to score the lesions, the observed dissociation is partly definitional. Please specify the expert-selection criterion and demonstrate a crossover interaction (syntax-lesioned model worse on syntax than semantics; semantic-lesioned model worse on semantics than syntax), not merely a general performance drop.
  2. [Full text (unavailable)] The submitted full text is corrupted and unreadable; no architecture details, lesion parameters, datasets, retraining protocol, or statistical tests can be checked. This is load-bearing because the abstract's empirical claims cannot be verified. A clean, complete manuscript must be provided for review.
  3. [Abstract] The claim that lesion outputs 'closely resemble' real patient speech is unsupported without quantitative details: no patient count, effect sizes, confidence intervals, or specification of the linguistic metrics (e.g., fluency, syntactic complexity, lexical diversity). Please report these values and, ideally, compare the two lesion models against each other and against control ablations using the same metrics.
  4. [Abstract] The rehabilitation result is stated as 'restores significant linguistic function' without any pre-post statistics, baseline, or control condition. It is unclear whether the improvement comes from retraining intact experts or merely from any fine-tuning on conversational data. Please include a no-retraining control and a control where damaged experts are not frozen, and report effect sizes with confidence intervals.
minor comments (2)
  1. [Abstract] The term 'healthy experts' is used without defining what 'healthy' means. Please clarify whether it refers to non-lesioned experts in the original model.
  2. [General] The title emphasizes 'Bridging Brains and Models,' but the abstract does not mention any direct neural data or brain measurements. Consider either adding such a connection or softening the title to match the actual content.

Circularity Check

0 steps flagged · score 0.0 of 10

No demonstrated circularity: external patient-speech validation and retraining benchmarks are independent of expert-selection; the unverified specialization premise is an auditability concern, not a self-definitional loop.

full rationale

The paper's claimed derivation chain is: identify functionally specialized MoE experts, lesion them, compare their linguistic outputs to real aphasic patient speech, then retrain the intact experts on conversational data. The abstract's key validation steps—comparison to real patient speech and retraining on conversational data—are external to whatever criterion was used to label experts as 'syntax' or 'semantics.' Therefore, even if expert selection were based on the same type of task used in evaluation, the claimed resemblance to Broca's and Wernicke's aphasia and the rehabilitation result do not reduce by construction to the selection inputs alone. The supplied full text is unreadable mojibake, so no equations, selection procedure, or self-citation chain can be audited. The abstract does assert 'functionally-specialized experts for syntax or semantics' without defining how that specialization was established; this is a genuine missing-support and auditability concern. However, under the hard rule requiring a quoted, specific reduction—such as an equation showing the predicted quantity equals the fitted input, or a parameter renamed as a prediction—this concern does not rise to a demonstrated circular step. No load-bearing self-citation is visible. Accordingly, the honest finding is no significant circularity, with the caveat that the unreadable full text prevents full verification of the expert-selection step.

Assumptions & free parameters 3 free parameters · 4 assumptions · 0 invented entities

Because only the abstract was readable, the ledger lists the premises the central claim must rely on rather than parameters I could measure. Free parameters are the expert-selection criterion, the lesion extent, and the retraining setup, all unspecified at the abstract level. The axioms are the domain assumptions that make the brain-to-model analogy work: clean functional specialization of experts, ablation as a surrogate for neural injury, the validity of the speech metrics, and retraining as an analogue of rehabilitation. No new physical or theoretical entities are introduced beyond the framing concept of a 'computational analogue.'

free parameters (3)
  • Expert selection criterion for syntax/semantics specialization
    How the model's experts are classified as syntax- or semantics-specialized (thresholds, probes, benchmarks) is not stated in the abstract; the lesion targets are chosen by this criterion, so it is a free choice that determines the observed deficit pattern.
  • Lesion extent (number/fraction of experts disabled)
    The severity of the simulated aphasia plausibly depends on how many experts are ablated; not specified in the abstract.
  • Retraining dataset and budget for rehabilitation
    The abstract says 'conversational data' was used for retraining but gives no dataset, size, or training schedule; the reported recovery magnitude depends on all of these.
assumptions (4)
  • domain assumption MoE experts are cleanly and functionally dissociable into syntax- and semantics-specialized modules.
    The entire lesion method presupposes that disabling specific experts produces a functionally selective deficit rather than a general degradation. This is the paper's core modeling premise, stated in the abstract ('functionally-specialized experts for syntax or semantics').
  • domain assumption Ablating a subset of an LLM's experts is a valid model of focal brain injury (functional lesion).
    The bridge from model to brain requires that component ablation in an artificial network meaningfully corresponds to neural damage; this homology is assumed throughout.
  • domain assumption The linguistic metrics used to compare model output with patient speech capture the clinically relevant dimensions of Broca's vs. Wernicke's aphasia.
    Validation 'against real patient speech' presupposes that the chosen comparison features (e.g., syntactic errors, fluency, paraphasias) are the right measurement axis; not substantiated in the abstract.
  • domain assumption Retraining intact experts on conversational text is an analogue of speech-language rehabilitation.
    The rehabilitation claim depends on the assumption that weight update on naturalistic text parallels therapeutic plasticity in human brains; no clinical validation is offered in the abstract.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Bridging Brains and Models: MoE-Based Functional Lesions for Simulating and Rehabilitating Aphasia." pith.science (2026). https://pith.science/paper/N7BI6TPA

@misc{pith2026250804749,
  author       = {Pith},
  title        = {Pith review of: Bridging Brains and Models: MoE-Based Functional Lesions for Simulating and Rehabilitating Aphasia},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/N7BI6TPA}},
  note         = {Machine review of arXiv:2508.04749}
}
read the original abstract

The striking alignment between large language models (LLMs) and human brain activity positions them as powerful models of healthy cognition. This parallel raises a fundamental question: if LLMs can model the intact brain, can we lesion them to simulate the linguistic deficits of the injured brain? In this work, we introduce a methodology to model aphasia - a complex language disorder caused by neural injury - by selectively disabling components in a modular Mixture-of-Experts (MoE) language model. We simulate distinct aphasia subtypes, validate their linguistic outputs against real patient speech, and then investigate functional recovery by retraining the model's remaining healthy experts. Our results demonstrate that lesioning functionally-specialized experts for syntax or semantics induces distinct impairments that closely resemble Broca's and Wernicke's aphasia, respectively. Crucially, we show that freezing the damaged experts and retraining the intact ones on conversational data restores significant linguistic function, demonstrating a computational analogue for rehabilitation. These findings establish modular LLMs as a powerful and clinically-relevant potential framework for modeling the mechanisms of language disorders and for computationally exploring novel pathways for therapy.

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

40 extracted references · 28 canonical work pages

  1. [1]

    , " * write output.state after.block = add.period write newline

    ENTRY address archivePrefix author booktitle chapter edition editor eid eprint howpublished institution isbn journal key month note number organization pages publisher school series title type volume year label extra.label sort.label short.list INTEGERS output.state before.all mid.sentence after.sentence after.block FUNCTION init.state.consts #0 'before.a...

  2. [2]

    write newline

    " write newline "" before.all 'output.state := FUNCTION n.dashify 't := "" t empty not t #1 #1 substring "-" = t #1 #2 substring "--" = not "--" * t #2 global.max substring 't := t #1 #1 substring "-" = "-" * t #2 global.max substring 't := while if t #1 #1 substring * t #2 global.max substring 't := if while FUNCTION word.in bbl.in capitalize " " * FUNCT...

  3. [3]

    Agirre, E.; Banea, C.; Cer, D.; Diab, M.; Gonzalez-Agirre, A.; Mihalcea, R.; Rigau, G.; and Wiebe, J. 2016. Semeval-2016 task 1: Semantic textual similarity, monolingual and cross-lingual evaluation. In Proceedings of the 10th international workshop on semantic evaluation (SemEval-2016), 497--511

    work page 2016

  4. [4]

    A.; Yeo, B

    Bertolero, M. A.; Yeo, B. T.; and D’Esposito, M. 2015. The modular and integrative functional architecture of the human brain. Proceedings of the National Academy of Sciences, 112(49): E6798--E6807

    work page 2015

  5. [5]

    K.; Teasell, R.; and Speechley, M

    Bhogal, S. K.; Teasell, R.; and Speechley, M. 2003. Intensity of aphasia therapy, impact on recovery. Stroke, 34(4): 987--993

    work page 2003

  6. [6]

    Caucheteux, C.; and King, J.-R. 2022. Brains and algorithms partially converge in natural language processing. Communications biology, 5(1): 134

    work page 2022

  7. [7]

    Cer, D.; Diab, M.; Agirre, E.; Lopez-Gazpio, I.; and Specia, L. 2017. Semeval-2017 task 1: Semantic textual similarity-multilingual and cross-lingual focused evaluation. arXiv preprint arXiv:1708.00055

    arXiv 2017

  8. [8]

    Chen, T.; Huang, S.; Xie, Y.; Jiao, B.; Jiang, D.; Zhou, H.; Li, J.; and Wei, F. 2022. Task-specific expert pruning for sparse mixture-of-experts. arXiv preprint arXiv:2206.00277

    arXiv 2022

Show all 40 references
  1. [9]

    Damasio, A. R. 1992. Aphasia. New England Journal of Medicine, 326(8): 531--539

    1992

  2. [10]

    S.; Schwartz, M

    Dell, G. S.; Schwartz, M. F.; Martin, N.; Saffran, E. M.; and Gagnon, D. A. 1997. Lexical access in aphasic and nonaphasic speakers. Psychological review, 104(4): 801

    1997

  3. [11]

    F.; and Ivanova, M

    Dronkers, N. F.; and Ivanova, M. V. 2023. The neuroscience of language and aphasia

    2023

  4. [12]

    F.; Wilkins, D

    Dronkers, N. F.; Wilkins, D. P.; Van Valin Jr, R. D.; Redfern, B. B.; and Jaeger, J. J. 2004. Lesion analysis of the brain areas involved in language comprehension. Cognition, 92(1-2): 145--177

    2004

  5. [13]

    Farah, M. J. 1991. Cognitive neuropsychology: Patterns of co-occurrence among the associative agnosias: Implications for visual object representation. Cognitive Neuropsychology, 8(1): 1--19

    1991

  6. [14]

    Fedus, W.; Zoph, B.; and Shazeer, N. 2022. Switch transformers: Scaling to trillion parameter models with simple and efficient sparsity. Journal of Machine Learning Research, 23(120): 1--39

    2022

  7. [15]

    Friederici, A. D. 2011. The brain basis of language processing: from structure to function. Physiological reviews, 91(4): 1357--1392

    2011

  8. [16]

    Gleichgerrcht, E.; Kocher, M.; Nesland, T.; Rorden, C.; Fridriksson, J.; and Bonilha, L. 2015. Preservation of structural brain network hubs is associated with less severe post-stroke aphasia. Restorative neurology and neuroscience, 34(1): 19--28

    2015

  9. [17]

    Glorot, X.; and Bengio, Y. 2010. Understanding the difficulty of training deep feedforward neural networks. In Proceedings of the thirteenth international conference on artificial intelligence and statistics, 249--256. JMLR Workshop and Conference Proceedings

    2010

  10. [18]

    A.; Zada, Z.; Ham, E.; Feder, A.; et al

    Goldstein, A.; Grinstein-Dabush, A.; Schain, M.; Wang, H.; Hong, Z.; Aubrey, B.; Nastase, S. A.; Zada, Z.; Ham, E.; Feder, A.; et al. 2024. Alignment of brain embeddings and artificial contextual embeddings in natural language points to common geometric patterns. Nature commun...

    2024

  11. [19]

    E.; and Shallice, T

    Hinton, G. E.; and Shallice, T. 1991. Lesioning an attractor network: investigations of acquired dyslexia. Psychological review, 98(1): 74

    1991

  12. [20]

    Q.; Sablayrolles, A.; Roux, A.; Mensch, A.; Savary, B.; Bamford, C.; Chaplot, D

    Jiang, A. Q.; Sablayrolles, A.; Roux, A.; Mensch, A.; Savary, B.; Bamford, C.; Chaplot, D. S.; Casas, D. d. l.; Hanna, E. B.; Bressand, F.; et al. 2024. Mixtral of experts. arXiv preprint arXiv:2401.04088

    2024 arXiv

  13. [21]

    F.; and Seidenberg, M

    Joanisse, M. F.; and Seidenberg, M. S. 1999. Impairments in verb morphology after brain injury: A connectionist model. Proceedings of the National Academy of Sciences, 96(13): 7592--7597

    1999

  14. [22]

    Kiran, S.; and Thompson, C. K. 2019. Neuroplasticity of language networks in aphasia: Advances, updates, and future challenges. Frontiers in neurology, 10: 295

    2019

  15. [23]

    Li, Y.; Su, H.; Shen, X.; Li, W.; Cao, Z.; and Niu, S. 2017. Dailydialog: A manually labelled multi-turn dialogue dataset. arXiv preprint arXiv:1710.03957

    2017 arXiv

  16. [24]

    Lichtheim, L.; et al. 1885. On aphasia. Broca’s Region, 318--347

  17. [25]

    MacWhinney, B.; Fromm, D.; Forbes, M.; and Holland, A. 2011. AphasiaBank: Methods for studying discourse. Aphasiology, 25(11): 1286--1307

    2011

  18. [26]

    B.; Islam, K

    Manir, S. B.; Islam, K. S.; Madiraju, P.; and Deshpande, P. 2024. Llm-based text prediction and question answer models for aphasia speech. IEEE Access

    2024

  19. [27]

    Marelli, M.; Bentivogli, L.; Baroni, M.; Bernardi, R.; Menini, S.; and Zamparelli, R. 2014. Semeval-2014 task 1: Evaluation of compositional distributional semantic models on full sentences through semantic relatedness and textual entailment. In Proceedings of the 8th internat...

    2014

  20. [28]

    Meunier, D.; Lambiotte, R.; and Bullmore, E. T. 2010. Modular and hierarchically modular organization of brain networks. Frontiers in neuroscience, 4: 200

    2010

  21. [29]

    Mou, X.; Ding, X.; He, Q.; Wang, L.; Liang, J.; Zhang, X.; Sun, L.; Lin, J.; Zhou, J.; Huang, X.; et al. 2024. From individual to society: A survey on social simulation driven by large language model-based agents. arXiv preprint arXiv:2412.03563

    2024 arXiv

  22. [30]

    Muennighoff, N.; Soldaini, L.; Groeneveld, D.; Lo, K.; Morrison, J.; Min, S.; Shi, W.; Walsh, P.; Tafjord, O.; Lambert, N.; et al. 2024. Olmoe: Open mixture-of-experts language models. arXiv preprint arXiv:2409.02060

    2024 arXiv

  23. [31]

    S.; O'Brien, J.; Cai, C

    Park, J. S.; O'Brien, J.; Cai, C. J.; Morris, M. R.; Liang, P.; and Bernstein, M. S. 2023. Generative agents: Interactive simulacra of human behavior. In Proceedings of the 36th Annual ACM Symposium on User Interface Software and Technology, 1--22. ACM

    2023

  24. [32]

    A.; Tuckute, G.; Kauf, C.; Hosseini, E

    Schrimpf, M.; Blank, I. A.; Tuckute, G.; Kauf, C.; Hosseini, E. A.; Kanwisher, N.; Tenenbaum, J. B.; and Fedorenko, E. 2021. The neural architecture of language: Integrative modeling converges on predictive processing. Proceedings of the National Academy of Sciences, 118(45): ...

    2021

  25. [33]

    Shazeer, N.; Mirhoseini, A.; Maziarz, K.; Davis, A.; Le, Q.; Hinton, G.; and Dean, J. 2017. Outrageously large neural networks: The sparsely-gated mixture-of-experts layer. arXiv preprint arXiv:1701.06538

    2017 arXiv

  26. [34]

    D.; Geranmayeh, F.; and Lambon Ralph, M

    Stefaniak, J. D.; Geranmayeh, F.; and Lambon Ralph, M. A. 2022. The multidimensional nature of aphasia recovery post-stroke. Brain, 145(4): 1354--1367

    2022

  27. [35]

    S.; and Karmiloff-Smith, A

    Thomas, M. S.; and Karmiloff-Smith, A. 2002. Modelling typical and atypical cognitive development. Handbook of childhood development, 575--599

    2002

  28. [36]

    Toneva, M.; and Wehbe, L. 2019. Interpreting and improving natural-language processing (in machines) with natural language-processing (in the brain). Advances in neural information processing systems, 32

    2019

  29. [37]

    Wang, C.; Fan, Z.; Han, Z.; Bi, Y.; and Li, J. 2025. Emergent modularity in large language models: Insights from aphasia simulations. bioRxiv, 2025--02

    2025

  30. [38]

    Warstadt, A.; Parrish, A.; Liu, H.; Mohananey, A.; Peng, W.; Wang, S.-F.; and Bowman, S. R. 2020. BLiMP: The benchmark of linguistic minimal pairs for English. Transactions of the Association for Computational Linguistics, 8: 377--392

    2020

  31. [39]

    M.; and Schneck, S

    Wilson, S. M.; and Schneck, S. M. 2020. Neuroplasticity in post-stroke aphasia: A systematic review and meta-analysis of functional imaging studies of reorganization of language processing. Neurobiology of Language, 2(1): 22--82

    2020

  32. [40]

    Zhang, Z.; Zeng, Z.; Lin, Y.; Xiao, C.; Wang, X.; Han, X.; Liu, Z.; Xie, R.; Sun, M.; and Zhou, J. 2023. Emergent modularity in pre-trained transformers. arXiv preprint arXiv:2305.18390

    2023 arXiv

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.