Language Impairment as a Lens onto Large Language Models

Muxuan He; Wang Bill Zhu; Yue Chen; Elsi Kaiser; Weizhi Du; Yao Du; Samuel Nastase; Yanjun Gao

cs.CLOct 20, 2026

Language Impairment as a Lens onto Large Language Models

Authors: Muxuan He, Wang Bill Zhu, Yue Chen, Elsi Kaiser, Weizhi Du, Yao Du, Samuel Nastase, Yanjun Gao

Organizations: University of Southern California · University of California, Berkeley · Independent Researcher · University of Colorado Anschutz Medical Campus

Abstract

Although large language models (LLMs) encode language in highly distributed spaces, they remain vulnerable to focal disruptions---a phenomenon strikingly similar to how localized brain damage causes clinical aphasia in humans. To investigate this theoretical tension, we ask whether LLMs possess a functional structure analogous to the human language system, by introducing a novel mechanistic interpretability framework inspired by aphasia. Through two complementary approaches, classifier-guided lesioning and continuous activation steering, we systematically induce aphasia-like breakdowns in LLMs. We find that language functions can be successfully isolated to discrete structural circuits and continuous representational vectors, yet these functions remain densely entangled with significant circuit overlap and vector similarities across distinct aphasic profiles. Furthermore, while steering generalizes robustly across downstream tasks, structural lesioning remains asymmetrically constrained. By bridging neuropsychology and mechanistic interpretability, we demonstrate that LLMs, functionally analogous to biological brains, rely on a densely entangled architecture that can be systematically mapped through clinically-informed interventions.

Figures & tables

Figure 1: Left : knocking out attention heads inside Llama-2-7B-chat makes its Cinderella story telling fragmented and incoherent , as in Broca’s aphasia; Right : adding a direction specific to Wernicke’s aphasia to the activations of the Qwen3.5-9B makes its story fluent but irrelevant , as in Wernicke’s aphasia.

	Cinderella		Sandwich
Group	Transcripts	Speakers	Transcripts	Speakers
Control	$395$	$288$	$472$	$360$
Anomic	$272$	$148$	$187$	$131$
Conduction	$141$	$73$	$81$	$62$
Broca	$273$	$113$	$111$	$77$
Wernicke	$58$	$30$	$29$	$23$

Table 1: Dataset statistics for the two AphasiaBank tasks.

Figure 2: Five-way classifier trained on AphasiaBank transcripts and labeled model scrambles using 14 text features and a text embedding, ensembled over 20 seeds.

Figure 3: Left : CLAMP ranks heads by their score alone, knocks out the top $k$ until the target probability exceeds 0.5, and prunes the set. Right : a subtype vector, the patients’ minus the controls’ mean hidden state, is added at a layer and strength chosen by the classifier.

Guided by corpus statistics
	Target
Method	Aph.	Broca	Wern.	All
Param	$2$	$0$	$0$	$2$
Param-mask	$4$	$0$	$0$	$4$
Param-mask-hard	$3$	$1$	$0$	$4$
Head-mask	$5$	$0$	$0$	$5$

Table 2: Number of settings in which a lesion method finds a lesion that makes the model’s answer reach the target, by the five-way classifier in 42 settings.

Figure 4: CLAMP lesions in 11 models, two tasks, and three targets: aphasia (Aph.), Broca’s and Wernicke’s aphasia. A cell gives the lesion’s number of heads, and its color the classifier’s target probability for the lesioned output. Gray cells mark no effect of lesioning.

Family	Features	Short Answer	MMLU
	TAB
GPT-2	$0.8\rightarrow 2.6$	$8/15$	$0.24$
Llama-2-chat	$0.0\rightarrow 1.8$	$3/15$	$1.59$
Pythia	$0.0\rightarrow 2.0$	$4/15$	$0.11$
Qwen3.5	$0.2\rightarrow 4.8$	$4.5/15$	$1.11$
All	$0.0\rightarrow 2.7$	$4/15$	$0.89$

Table 3: What the CLAMP lesions do beyond the narrative they were found on, as medians over the lesions that reach their target. Features : aphasia features of the TAB, of 19, that the LLM judge finds per connected-text answer of the intact and of the lesioned model. Short Answer : TAB short-answer items for comprehension and repetition that the lesioned model gets right, of 15. MMLU : drop in the log-probability of the correct answer on two MMLU subjects, in nats.

Figure 5: Expert labels of the answers steered toward Broca’s or Wernicke’s aphasia, by target and prompt. Source task : the narrative whose answers chose the steering layer and strength; other prompts : the other narrative and the five connected-text prompts of the Text Aphasia Battery (TAB). An answer counts as reaching the target when either expert labels it so; any other answer is split evenly between its two experts’ labels.

	Two-way	Five-way	All
Classifier	$\boldsymbol{73}$	$\boldsymbol{62}$	$\boldsymbol{65}$
LLM judge	$64$	$51$	$55$
Two experts	$88$	$76$	$80$
Cohen’s $\kappa$	$0.75$	$0.68$	$0.76$

Table 4: Agreement (%) between automated labelers and two human experts on steered model responses. Two-way: Binary classification (aphasic vs. non-aphasic) of responses steered toward aphasia. Five-way: Five-class classification of responses steered toward Broca’s or Wernicke’s aphasia.

Appendix figures & tables34 assets

Supplementary material from the paper’s appendix.

Appendix

GPT-2 ( Radford et al., 2019 ) : base model, question-and-answer prompts
Model	Hugging Face identifier	Layers	Heads per layer	Heads in all	Softmax heads
GPT-2 XL	gpt2-xl	$48$	$25$	$1{,}200$	$1{,}200$
Llama-2 ( Touvron et al., 2023 ) : chat models, Llama-2 chat template
Llama-2-7B-chat	meta-llama/Llama-2-7b-chat-hf	$32$	$32$	$1{,}024$	$1{,}024$
Llama-2-13B-chat	meta-llama/Llama-2-13b-chat-hf	$40$	$40$	$1{,}600$	$1{,}600$
Pythia ( Biderman et al., 2023 ) : base models, question-and-answer prompts

Appendix

Table 5: The 11 models, grouped by family, with the kind of model and the prompt format of each family. In Qwen3.5, every fourth layer, layers 3, 7, 11, and so on when counting from 0, uses softmax attention, and the other layers use Gated DeltaNet linear attention. For Qwen3.5, Heads per layer gives the heads of a softmax layer and of a Gated DeltaNet layer, and a lesion can knock out heads of both kinds. Softmax heads counts the heads of the softmax layers, which in the other models are all of the heads.

Figure 6: t-SNE of each task’s transcripts and of the classifier’s model scrambles, in the text space of the classifier’s encoder, BGE-large, without the classifier’s fine-tuning. The maps hold the 1,139 Cinderella and 880 sandwich transcripts of Table 1 , with 169 and 90 model scrambles. Conduction and anomic transcripts share one color because they share one class of the classifier.

Family	Prompt template
GPT-2 XL and Pythia	Question: request \n Answer:
Llama-2-chat	[INST] request [/INST]
Qwen3.5	<\|im_start\|>user\n request <\|im_end\|>\n <\|im_start\|>assistant\n <think>\n\n</think>\n\n
Request	Text
Yes-or-no question	Do you know the story of Cinderella?
Cinderella Story request	Tell me as much of the story of Cinderella as you can.

Appendix

Table 6: The prompts, verbatim. Each prompt is the template of the model’s family with one of the three requests in place of request , and \n marks a newline. The prompts of GPT-2 XL and of the Llama-2 models start with a beginning-of-sequence token, <|endoftext|> or <s> , which the templates leave out, and the prompts of Pythia and Qwen3.5 start with no such token. Neither chat template adds a system prompt, and the Qwen3.5 template, with thinking turned off, closes an empty thinking block, so that the answer starts at once.

Subtest	Prompt	Target
Connected text	Tell me about the best trip you ever took. Describe a happy childhood memory. Tell me about your first job. What do you like about where you live? Describe the steps to make a simple meal.	–
Word comprehension	Which one is an animal with a mane? (lion, drum, violin, giraffe, boot, boat)	lion
Sentence comprehension	Are babies watched by babysitters?	yes
Repetition	Please repeat exactly: The sun rises in the East.	The sun rises in the East.

Appendix

Table 7: The items of the Text Aphasia Battery (TAB) on which we test every lesion, transcribed from the TAB paper ( Roll et al., 2026b ) : all five connected-text prompts and one of the five items of each other subtest, verbatim. Connected-text prompts follow the TAB’s instruction Respond to the following prompt in 3-5 full sentences: , and the other items follow the instruction You are having a conversation. Answer the question directly and briefly. The five-way classifier and the LLM judge score the connected-text answers, and rules compare the other answers with their targets.

AphasiaBank transcripts
	Cinderella	Sandwich	Total
Control	$399$	$936$	$1{,}335$
Conduction or anomic	$417$	$334$	$751$
Anomic	$276$	$235$	$511$
Conduction	$141$	$99$	$240$
Broca	$273$	$131$	$404$

Appendix

Table 8: Training texts of the five-way classifier, by class and task prompt. The AphasiaBank transcripts carry six group labels, and the classifier merges anomic and conduction aphasia into one class. The model scrambles are answers of Llama-2-7B-chat and GPT-2 XL from a preliminary knockout study. Distinct texts counts each text once, because 557 sandwich texts and 4 Cinderella texts appear twice among the transcripts.

Narrative measures
Feature	Definition
Utterances	Number of utterances, each ending in “.”, “!”, or “?”
Word tokens	Number of words
Mean length of utterance	Mean number of words per utterance, leaving out utterances that are a single filler word such as “uh” or “yeah”
Repetition statistics
Shortest repeated unit	Length in words of the shortest repeated unit, or 0 if there is none

Appendix

Table 9: The 14 features that the five-way classifier reads besides its text encoder, each standardized with the mean and standard deviation of the training split. A repeated unit is a sequence of 1 to 100 tokens, words or punctuation marks, that occurs several times in a row, or a sequence of six or more tokens that occurs again anywhere in the text. Similarities are mean cosine similarities between the all-MiniLM-L6-v2 embedding of the text and those of the human transcripts of a group, from both tasks. Because of a label mismatch in the code, the last feature subtracts zero instead of the similarity to model scrambles.

All 2,798 held-out texts
Class	Precision	Recall	F1	N
Control	$0.942$	$0.981$	$0.961$	$1{,}317$
Cond. or anomic	$0.828$	$0.769$	$0.797$	$737$
Broca	$0.783$	$0.816$	$0.799$	$397$
Wernicke	$0.583$	$0.457$	$0.512$	$92$
Model scrambles	$0.992$	$0.996$	$0.994$	$255$

Appendix

Table 10: Precision, recall, and F1 of the five-way classifier on held-out texts, by class. Each of 20 models holds out a random 20% of the 2,843 training texts, and each held-out text is scored by the mean output of the models that held it out. The lower block keeps the held-out texts that occur only once in the training data, so that no copy of them can be in a training split. Cond. or anomic: conduction or anomic aphasia; N: number of texts.

	Predicted class
		Cond. or
True class	Control	anomic	Broca	Wern.	MS
Control	$\boldsymbol{1{,}292}$	$25$	$0$	$0$	$0$
Cond. or anomic	$78$	$\boldsymbol{567}$	$75$	$17$	$0$
Broca	$0$	$58$	$\boldsymbol{324}$	$13$	$2$
Wernicke	$1$	$35$	$14$	$\boldsymbol{42}$	$0$

Appendix

Table 11: Confusion matrix of the five-way classifier on the 2,798 held-out texts, with true classes as rows and predicted classes as columns; bold counts are correct predictions. Each held-out text is scored by the mean output of the models, of 20, whose random test split holds it. Cond. or anomic: conduction or anomic aphasia; Wern.: Wernicke’s aphasia; MS: model scrambles.

		Target
Model	Task	Aph.	Broca	Wern.
Llama-2-13B-chat	Cinderella	$1{,}305$	$1{,}307$	$1{,}313$
	Sandwich	$1{,}298$	$1{,}298$	$1{,}298$
Llama-2-7B-chat	Cinderella	$568$	$569$	$571$
	Sandwich	$449$	$450$	$452$
GPT-2 XL	Cinderella	$569$	$569$	$569$

Appendix

Table 12: Heads that tie in the single-head ranking of CLAMP : the number of heads whose knockout alone leaves the target probability within

10^{-6}

of the intact model’s value, by model, task, and target. The models have, from top to bottom, 1,600, 1,024, 1,200, 1,024, and 128 heads. Tied heads keep their layer order in the ranking, so a growth that reaches them adds them from the first layer upward.

Heads
	Target				Models
Lesion space and search	Aph.	Broca	Wern.	All	GPT-2 XL	Llama-2-chat	Pythia
Head-mask	$5$	$0$	$0$	$5$	$1$	$0$	$4$
Head-mask-hard	$8$	$1$	$1$	$10$	$1$	$0$	$9$
Head-grown-cos	$2$	$0$	$0$	$2$	$1$	$0$	$1$
Layer search, first stage	$5$	$1$	$2$	$8$	$0$	$0$	$8$

Appendix

Table 13: Number of settings in which each lesion method finds a lesion that works, that is, a lesion whose answer counts and reaches the target: a target probability above 0.5 by the five-way classifier, or the target as its most probable class, for every lesion space and search that we ran on the seven older models. The 42 settings combine seven models, two tasks, and three targets, so each target has 14 settings, GPT-2 XL has 6, the two Llama-2-chat models have 12, and the four Pythia models have 24. An MLP or attention lesion zeroes the top 1% of the weights of one MLP or attention weight matrix by

|w\,\partial L/\partial w|

on control transcripts, and an MLP block is a block of 256 neurons of one MLP layer. Within each space, the searches guided by corpus statistics come before those guided by the classifier. Grown : lesions ranked by their single-lesion target probability, grown in that order as CLAMP grows heads, and pruned. Then head sweep and Param-C : when the grown MLP lesions do not work, heads or, for Param-C , attention lesions are added layer by layer on top of them. Layer search : every head alone, then the top heads of each layer together, and in later stages heads and MLP blocks added layer by layer. Bold marks the largest count in each column.

Heads
						Length		Held out
Lesion space and search	Sets	Size	Target	Scr.	zlib	Chars	${<}100$	TAB	Other	Degen.
Head-mask	$5$	$97\phantom{.5}$	$0.74$	$0.000$	$0.72$	$199$	$1$	$0.114$	$0.165$	$0$
Head-mask-hard	$10$	$9\phantom{.5}$	$0.85$	$0.001$	$0.80$	$135$	$4$	$0.035$	$0.009$	$0$
Head-grown-cos	$2$	$22.5$	$0.74$	$0.256$	$0.48$	$470$	$0$	$0.009$	$0.008$	$1$
Layer search, first stage	$8$	$2.5$	$0.67$	$0.015$	$0.80$	$158$	$2$	$0.012$	$0.007$	$0$

Appendix

Table 14: The working sets of each lesion method of Table 13 on the seven older models, one set per setting in which the method works, described by medians except in the count columns. Sets : number of working sets. Size : components in the set, that is, heads, lesions, or MLP blocks, where a soft mask counts the components with an amount above 0.0005. Target : target probability of the lesioned model’s answer. Scr. :

P(\mathrm{scrambles})

, the classifier’s probability of model scrambles for that answer. zlib : length of the zlib-compressed answer over the length of the answer, both in bytes, which falls below 0.25 for an answer that repeats itself and exceeds 1 for some short answers. Chars : answer length in characters, not counting Unicode format characters;

{<}100

: number of answers shorter than 100 characters. TAB : mean target probability over the five TAB connected-text answers with the set knocked out. Other : target probability of the answer to the other task with the set knocked out, over the sets whose answer there counts. Degen. : number of degenerate answers, which have a zlib ratio below 0.25 or model scrambles as the most likely class. The results of Param store neither held-out check.

Model	Task	Target	Heads	$P$	Intact answer	Lesioned answer
GPT-2 XL	Cinderella	Aphasia	$9$	$0.891$	…Cinderella was born in a small village called "Horse-head" (or Horse-Head) in the Kingdom of Khazaria.…	I don’t know anything about her, but she’s a pretty girl and she has to go to the ball.
Pythia-1B	sandwich	Aphasia	$15$	$0.691$	I’d use the same recipe as for my bread, but add some of your favorite jam or preserves to it instead of mayonnaise.…	You’d put the bread in first, then add some of that to your sandwich. Then you could just eat it all!
Qwen3.5-2B	sandwich	Broca	$18$	$0.738$	Making the perfect Peanut Butter & Jelly (PB&J) is less about following strict rules…	Tell, tell the way of? / "Tell to know." "
Pythia-6.9B	sandwich	Broca	$13$	$0.665$	I’d put the bread on top of the PB, then spread it with some jam (or honey)…	You have to ask for it!
Llama-2-13B-chat †	Cinderella	Broca	$101$	$0.886$	Sure! Here’s a brief retelling of the classic tale of Cinderella: / Once upon a time, there was a beautiful and kind girl named Cinderella…	to tell a tale of Ct.. to tell me.
Llama-2-13B-chat	Cinderella	Wernicke	$109$	$0.786$	as in the row above	tours and tours, but I can’dough, but i can be a new to the whole and the previous one is not.

Appendix

Table 15: Answers under CLAMP lesions that reach their target, two for each target from different model families, and the Llama-2-13B-chat example of § 5.1 , marked † . Target : aphasia, Broca’s aphasia, or Wernicke’s aphasia; Heads : the number of heads that the lesion knocks out;

P

: the five-way classifier’s probability of the target for the lesioned answer, which is at most 0.009 for each intact answer. The classifier’s most likely class for each intact answer is control speech, except for the GPT-2 XL story, which it calls a model scramble. Each answer is the model’s single greedy answer; excerpts are verbatim, … marks omitted text, and a slash marks one or more line breaks.

Model	Task	Lesion	$P(\mathrm{scrambles})$	Lesioned answer
Llama-2-7B-chat	Cinderella	Heads L0H25, L1H8, and L1H29	$0.9988$	I’m not sure if you could provide more information about your company, such as its website or any other relevant details, please let me know." / I understand that is a polite
Pythia-1.4B	Cinderella	Param cluster for Wernicke’s aphasia, 23 weight lesions	$0.9999$	The first time I was a Cinderella, it is in Cindere-lta, by Cindere-tant, T he first t e r s A c n d i l -n t h E D L O N S U P R E M B L Y T H E F O R T W E T H E F O R T W E…
Pythia-6.9B	sandwich	Param cluster for Broca’s aphasia, 32 weight lesions	$0.9981$	Click to,vumo-click here for more information on / Click Here To Buy Now!

Appendix

Table 16: Answers under lesions found by criteria that cannot recognize a machine failure, all of which the five-way classifier calls model scrambles. The three heads of Llama-2-7B-chat, where L

i

j

is head

j

of layer

i

, come from the preliminary search of § 3 , whose classifier had only an aphasic and a control class. Each weight lesion of Param zeroes the 1% of one weight matrix’s weights with the largest

|w\,\partial L/\partial w|

(§ 4.1 ), and each Param cluster reaches its aphasia type by Param ’s own criterion, BLEU-1 and BERTScore against patient transcripts.

P(\mathrm{scrambles})

: the five-way classifier’s probability of model scrambles, to four decimals. Answers are verbatim and complete except where … marks omitted text, and a slash marks one or more line breaks; the Pythia-1.4B answer goes on repeating “ T H E F O R T W E ” until it ends.

				Narrative		TAB			MMLU
Model	Task	Target	Heads	Same	Other	Short Answer	Features	Impaired	Drop
GPT-2 XL	Cinderella	Aphasia	$9$	$0.891$	$0.009$	$10/15$	$0.8\rightarrow 2.6$	$0.4\rightarrow 0.6$	$0.08$
	Cinderella	Wernicke	$244$	$0.647$	$0.000$	$4/15$	$0.8\rightarrow 4.4$	$0.4\rightarrow 0.8$	$0.21$
	Sandwich	Aphasia	$31$	$0.862$	$0.859$	$9/15$	$0.8\rightarrow 0.6$	$0.4\rightarrow 0.2$	$0.27$
	Sandwich	Broca	$146$	$0.884$	$0.015$	$4/15$	$0.8\rightarrow 3.4$	$0.4\rightarrow 0.8$	$0.36$
	Sandwich	Wernicke	$188$	$0.717$	$0.001$	$8/15$	$0.8\rightarrow 2.0$	$0.4\rightarrow 0.4$	$0.24$

Appendix

Table 17: The CLAMP lesions of the base models GPT-2 XL and Pythia that reach their target, tested beyond the narrative task that they were found on. Heads : number of heads knocked out. Same and Other : the five-way classifier’s probability of the target for the lesioned model’s answer to the narrative task that the lesion was found on and to the other narrative task. Short Answer : short-answer items of the Text Aphasia Battery (TAB) for comprehension and repetition that the lesioned model answers correctly, of 15; the intact models answer 9 (GPT-2 XL), 10 (Pythia-1B), 8 (Pythia-1.4B), 7 (Pythia-2.8B), and 6 (Pythia-6.9B) correctly. Features and Impaired : the mean number of aphasia features, of 19, that the LLM judge finds per connected-text answer, and the share of the five answers that it judges to show an overall communication impairment, for the intact model before the arrow and for the lesioned model after it. Drop : drop in the mean log-probability of the correct MMLU answer letter, in nats, averaged over two subjects.

				Narrative		TAB			MMLU
Model	Task	Target	Heads	Same	Other	Short Answer	Features	Impaired	Drop
Llama-2-7B-chat	Cinderella	Aphasia	$23$	$0.606$	$0.001$	$2/15$	$0.0\rightarrow 0.6$	$0.0\rightarrow 0.0$	$2.91$
	Cinderella	Broca	$42$	$0.678$	$0.000$	$7/15$	$0.0\rightarrow 3.6$	$0.0\rightarrow 1.0$	$1.08$
	Cinderella	Wernicke	$276$	$0.682$	$0.019$	$0/15$	$0.0\rightarrow 3.0$	$0.0\rightarrow 0.6$	$3.59$
	Sandwich	Aphasia	$60$	$0.931$	$0.000$	$3/15$	$0.0\rightarrow 2.2$	$0.0\rightarrow 0.8$	$1.14$
	Sandwich	Wernicke	$160$	$0.736$	$0.000$	$1/15$	$0.0\rightarrow 4.4$	$0.0\rightarrow 1.0$	$2.24$

Appendix

Table 18: The CLAMP lesions of the chat models Llama-2-chat and Qwen3.5 that reach their target, tested beyond the narrative task that they were found on. Heads : number of heads knocked out. Same and Other : the five-way classifier’s probability of the target for the lesioned model’s answer to the narrative task that the lesion was found on and to the other narrative task. Short Answer : short-answer items of the Text Aphasia Battery (TAB) for comprehension and repetition that the lesioned model answers correctly, of 15; the intact models answer 7 (Llama-2-7B-chat), 8 (Llama-2-13B-chat), 13 (Qwen3.5-0.8B and Qwen3.5-2B), and 14 (Qwen3.5-4B and Qwen3.5-9B) correctly. Features and Impaired : the mean number of aphasia features, of 19, that the LLM judge finds per connected-text answer, and the share of the five answers that it judges to show an overall communication impairment, for the intact model before the arrow and for the lesioned model after it. Drop : drop in the mean log-probability of the correct MMLU answer letter, in nats, averaged over two subjects. † The answer does not count, because the lesioned model denies knowing the story of Cinderella.

				Probability		Share of heads			TAB short answer
					Restore					Restore	Best
Model	Task	Target	Heads	Lesion	all	Restored	Left out	Alone	Lesion	all	restore
GPT-2 XL	Cinderella	Aphasia	$9$	$0.891$	$0.001$	$0.67$	$1.00$	$0.00$	$10/15$	$9/15$	$10/15$
		Wernicke	$244$	$0.647$	$0.001$	$0.75$	$0.95$	$0.00$	$4/15$	$7/15$	$5/15$
	Sandwich	Aphasia	$31$	$0.862$	$0.005$	$0.81$	$1.00$	$0.00$	$9/15$	$7/15$	$9/15$
		Broca	$146$	$0.884$	$0.000$	$0.91$	$1.00$	$0.00$	$4/15$	$9/15$	$5/15$

Appendix

Table 19: Activation patching of the CLAMP lesions of the base models GPT-2 XL and Pythia that reach their targets. Heads : the number of heads in the set. Probability : the five-way classifier’s probability of the target with the set knocked out, Lesion , and with the clean outputs of all of its heads written back while the model reads the prompt, Restore all , so that the set is knocked out only while the model writes its answer; an answer undoes the lesion when it no longer reaches the target, a probability of 0.5 or below with the target not the most probable class, and the one restore-all answer that still reaches it is in bold. Share of heads : the share of the set’s heads whose clean output on the prompt undoes the lesion, Restored , whose removal from the set undoes it, Left out , and that reach the target when knocked out alone, Alone . TAB short answer : short-answer items of the Text Aphasia Battery (TAB) for comprehension and repetition that the model answers correctly, of 15, with the set knocked out, with all heads restored, and with the best of at most eight single-head restores; the intact models answer 9 (GPT-2 XL), 10 (Pythia-1B), 8 (Pythia-1.4B), 7 (Pythia-2.8B), and 6 (Pythia-6.9B) correctly.

				Probability		Share of heads			TAB short answer
					Restore					Restore	Best
Model	Task	Target	Heads	Lesion	all	Restored	Left out	Alone	Lesion	all	restore
Llama-2-7B-chat	Cinderella	Aphasia	$23$	$0.606$	$0.012$	$0.61$	$1.00$	$0.00$	$2/15$	$7/15$	$9/15$
		Broca	$42$	$0.678$	$0.001$	$0.79$	$1.00$	$0.00$	$7/15$	$7/15$	$7/15$
		Wernicke	$276$	$0.682$	$0.001$	$0.88$	$0.99$	$0.00$	$0/15$	$5/15$	$1/15$
	Sandwich	Aphasia	$60$	$0.931$	$0.007$	$0.95$	$1.00$	$0.00$	$3/15$	$7/15$	$3/15$

Appendix

Table 20: Activation patching of the CLAMP lesions of the chat models Llama-2-chat and Qwen3.5 that reach their targets. Heads : the number of heads in the set. Probability : the five-way classifier’s probability of the target with the set knocked out, Lesion , and with the clean outputs of all of its heads written back while the model reads the prompt, Restore all , so that the set is knocked out only while the model writes its answer; an answer undoes the lesion when it no longer reaches the target, a probability of 0.5 or below with the target not the most probable class. Share of heads : the share of the set’s heads whose clean output on the prompt undoes the lesion, Restored , whose removal from the set undoes it, Left out , and that reach the target when knocked out alone, Alone . TAB short answer : short-answer items of the Text Aphasia Battery (TAB) for comprehension and repetition that the model answers correctly, of 15, with the set knocked out, with all heads restored, and with the best of at most eight single-head restores; the intact models answer 7 (Llama-2-7B-chat), 8 (Llama-2-13B-chat), 13 (Qwen3.5-0.8B and Qwen3.5-2B), and 14 (Qwen3.5-4B and Qwen3.5-9B) correctly.

GPT-2 XL
	Heads			Aphasia and Broca			Aphasia and Wernicke			Broca and Wernicke
Task	Aph.	Broca	Wern.	Shared	Chance	$p$	Shared	Chance	$p$	Shared	Chance	$p$
Cinderella	$9$	–	$244$	–	–	–	$0$	$1.83$	$1$	–	–	–
Sandwich	$31$	$146$	$188$	$18$	$3.77$	$<\!0.001$	$12$	$4.86$	$0.0014$	$70$	$22.87$	$<\!0.001$
Pythia-1B
Cinderella	$8$	$27$	$13$	$4$	$1.69$	$0.061$	$2$	$0.81$	$0.19$	$7$	$2.74$	$0.0062$

Appendix

Table 21: Heads shared by the CLAMP lesions that a model has for two targets on the same task, for every pair of targets whose two lesions reach their target. Heads : the number of heads in the lesion for the aphasia (Aph.), Broca, and Wernicke (Wern.) target, where a dash marks a target without a lesion. Shared : the heads that the two lesions of a pair share. Chance : the number of shared heads expected if both lesions were drawn at random from the model’s

N

heads,

{|L_{1}|\,|L_{2}|/N}

for lesions

L_{1}

and

L_{2}

p

: the one-sided hypergeometric probability of sharing at least as many heads by chance. All pairs : the total of shared heads, the total expected by chance, and the number of pairs with

p<0.01

. Median share : the median over pairs of the shared heads as a share of the smaller lesion. Three combinations of model and task with fewer than two lesions, Qwen3.5-0.8B on sandwich and Qwen3.5-9B on both tasks, are left out.

GPT-2 XL
	Heads						Union		Intersection
Target	Cinderella	Sandwich	Shared	Chance	Ratio	$p$	Cinderella	Sandwich	Cinderella	Sandwich
Aphasia	$9$	$31$	$1$	$0.23$	$4.30$	$0.21$	$0.000$	$0.001$	$0.015$	$0.007$
Wernicke	$244$	$188$	$92$	$38.23$	$2.41$	$1.9\times 10^{-22}$	$0.000$	$0.000$	$0.000$	$0.000$
Llama-2-13B-chat
Aphasia	$77$	$90$	$24$	$4.33$	$5.54$	$2.5\times 10^{-13}$	$0.001$	$0.003$	$0.018$	$0.001$

Appendix

Table 22: Heads shared by the CLAMP lesions that a model has for the same target on the Cinderella and on the sandwich task, in one model per family. Heads : the number of heads in the Cinderella set

C

, in the sandwich set

S

, and in both. Chance : the number of shared heads expected if both sets were drawn at random from the model’s

N

heads,

{|C|\,|S|/N}

. Ratio : shared heads divided by chance.

p

: the one-sided hypergeometric probability of sharing at least as many heads by chance. Union and Intersection : the classifier’s probability of the target for the answer to each task with the heads of either set, or only the shared heads, knocked out; values above 0.5, in bold, reach the target, and a dash marks pairs without shared heads. Three pairs are left out because a set does not reach its target: Broca’s aphasia in GPT-2 XL, whose Cinderella set does not reach it, and Broca’s and Wernicke’s aphasia in Qwen3.5-9B, whose sets reach it on neither task.

Cinderella task
		AUC, lowest to highest over layers				Severity
Model	Layers	Aphasia	Broca	Wernicke	Transfer	$\rho$	Layer	PC1
GPT-2 XL	$48$	$0.981$ – $0.993$	$0.944$ – $0.951$	$0.773$ – $0.851$	$0.871$ – $0.958$	$-0.759$	$44$	$0.955$
Llama-2-7B-chat	$32$	$0.980$ – $0.994$	$0.937$ – $0.947$	$0.764$ – $0.850$	$0.884$ – $0.971$	$-0.797$	$11$	$0.948$
Llama-2-13B-chat	$40$	$0.979$ – $0.995$	$0.935$ – $0.947$	$0.765$ – $0.857$	$0.889$ – $0.978$	$-0.806$	$17$	$0.943$
Pythia-1B	$16$	$0.982$ – $0.992$	$0.939$ – $0.949$	$0.772$ – $0.834$	$0.889$ – $0.970$	$-0.788$	$12$	$0.942$

Appendix

Table 23: Linear probes and directions in the residual stream that each model computes while it reads the human transcripts of a task as its own answer. Each AUC column gives the lowest and the highest area under the ROC curve over all layers. Aphasia : transcripts of the four aphasic groups against control transcripts, on speakers held out by five-fold cross-validation. Broca and Wernicke : transcripts of that type against all other transcripts, including those of the other types of aphasia, evaluated in the same way. Transfer : the aphasia probe fitted on all transcripts of the row’s task and tested on all transcripts of the other task. Severity : the most negative Spearman correlation

\rho

, over layers, between the projection of an aphasic transcript on the aphasia vector and the speaker’s aphasia quotient, where lower quotients mean more severe aphasia, and the layer at which it occurs, counted from 0. PC1 : the absolute cosine between the aphasia vector and the first principal component of all transcripts’ token-averaged residual streams, averaged over layers.

	Cinderella task							Sandwich task
	Aphasia		Broca		Wernicke			Aphasia		Broca		Wernicke
Model	$n$	$r$	$n$	$r$	$n$	$r$	Random	$n$	$r$	$n$	$r$	$n$	$r$	Random
GPT-2 XL	$2$	$2.75$	$3$	$4$	$0$	–	$5$	$4$	$1.25$	$4$	$2.25$	$0$	–	$2$
Llama-2-7B-chat	$5$	$2$	$3$	$3$	$0$	–	$1$	$4$	$2$	$1$	$3$	$0$	–	$1$
Llama-2-13B-chat	$3$	$2$	$3$	$3$	$0$	–	$0$	$3$	$3$	$4$	$3$	$0$	–	$3$
Pythia-1B	$8$	$0.5$	$6$	$1$	$4$	$0.75$	$8$	$8$	$1$	$8$	$1$	$1$	$0.5$	$12$

Appendix

Table 24: Steering sweep of each model’s answer to the task prompt with the aphasia, Broca, and Wernicke vectors of the same task, at 8 layers and 8 strengths from

r=0.1

4

, where

r

is the norm of the added vector relative to the mean norm of the transcripts’ token-averaged residual stream at that layer.

n

: the number of the 8 layers at which some strength makes the classifier’s probability of the vector’s target exceed 0.5.

r

: the median, over these layers, of the smallest such strength, or a dash where no layer works. Random : the number of the 192 settings, three random directions with the aphasia vector’s norm at the same 8 layers and 8 strengths, at which

P(\mathrm{aphasia})

exceeds 0.5. The last row sums

n

over the 11 models, of 88 layers per task, takes the median smallest strength over all working layers, and sums the random settings, of 2,112 per task.

Figure 7: $P(\mathrm{aphasia})$ of the answer to the task prompt when the aphasia vector of the same task is added at one of 8 layers, counted from 0, with one of 8 strengths $r$ , in the largest model of each family. The 8 strengths run from $r=0.1$ to $4$ , where $r$ is the norm of the added vector relative to the mean norm of the human transcripts’ token-averaged residual stream at that layer. The middle tick of the color bar, 0.5, is the threshold above which an answer reaches the aphasia target. A dot marks answers whose most probable class is model scrambles, and hatched cells mark answers shorter than 20 characters, which are not scored. The unsteered Cinderella answers of GPT-2 XL and Qwen3.5-9B already have model scrambles as their most probable class.

		Aphasia			Broca			Wernicke
Model	Source task	Layer	$r$	$P$	Layer	$r$	$P$	Layer	$r$	$P$
GPT-2 XL	Cinderella	$30$	$1.5$	$0.865$	$18$	$1$	$0.535$	$30$	$1$	$0.276^{\dagger}$
	Sandwich	$36$	$1$	$0.803$	$18$	$1$	$0.697$	$36$	$1$	$0.422^{\dagger}$
Llama-2-7B-chat	Cinderella	$0$	$1.5$	$0.994$	$12$	$3$	$0.957$	$16$	$3$	$0.011^{\dagger}$
	Sandwich	$0$	$1.5$	$0.659$	$12$	$3$	$0.995$	$24$	$3$	$0.046^{\dagger}$
Llama-2-13B-chat	Cinderella	$20$	$1.5$	$0.524$	$20$	$2$	$0.570$	$30$	$3$	$0.414^{\dagger}$

Appendix

Table 25: Steering settings of the human evaluation, for the seven models whose steered answers two experts labeled. For each model, source task, and target, the vector comes from the transcripts of the source task, and the setting is the smallest strength

r

at which the target probability of the steered answer to the source task exceeds 0.5 in the sweep, at the layer, counted from 0, with the highest probability at that strength.

P

: the classifier’s probability of the target for that answer. † No setting of the sweep passes 0.5, so the setting with the highest probability is used.

Figure 8: The models’ answers among the transcripts and model scrambles of Figure 6 , in one t-SNE per task fitted on all of them together. Each panel shows the intact answers of the 11 models as white circles, and either their answers lesioned by CLAMP toward Broca’s or Wernicke’s aphasia, as triangles in the top row, or their answers steered toward these types, as diamonds in the bottom row. A hollow marker is an answer that does not reach its target.

	Broca against Wernicke		Same type, across tasks		Shuffled labels			Type-specific parts
Model	Cinderella	Sandwich	Broca	Wernicke	Cinderella	Sandwich	Below	Cinderella	Sandwich
GPT-2 XL	$0.90$	$0.81$	$0.62$	$0.59$	$0.99$	$0.97$	$96/96$	$-0.47$	$-0.43$
Llama-2-7B-chat	$0.87$	$0.79$	$0.66$	$0.69$	$0.99$	$0.97$	$64/64$	$-0.48$	$-0.42$
Llama-2-13B-chat	$0.86$	$0.79$	$0.66$	$0.69$	$0.99$	$0.97$	$78/80$	$-0.49$	$-0.41$
Pythia-1B	$0.87$	$0.80$	$0.72$	$0.63$	$0.99$	$0.97$	$32/32$	$-0.65$	$-0.43$
Pythia-1.4B	$0.86$	$0.83$	$0.72$	$0.64$	$0.99$	$0.97$	$48/48$	$-0.63$	$-0.43$

Appendix

Table 26: Cosine similarity of steering vectors, as the median over a model’s layers. Broca against Wernicke : the Broca and the Wernicke vector of the same task. Same type, across tasks : the Cinderella and the sandwich vector of one type. Shuffled labels : the Broca and the Wernicke vector after shuffling the two labels among the transcripts of both types, keeping the group sizes, averaged over 200 shuffles. Below : layers at which the observed Broca-against-Wernicke cosine lies below the 2.5th percentile of the shuffled cosines, over both tasks. Type-specific parts : the cosine of the Broca and the Wernicke vector after projecting the aphasia vector out of both.

Against expert 1
	Two-way	Five-way	All
Classifier	$\boldsymbol{64}$	$\boldsymbol{48}$	$\boldsymbol{53}$
Cohen’s $\kappa$	$0.25$	$0.33$	$0.44$
LLM judge	$62$	$41$	$48$
Cohen’s $\kappa$	$0.22$	$0.21$	$0.38$
Against expert 2

Appendix

Table 27: Agreement of the classifier and the LLM judge with each of the two experts and with the experts’ consensus, in percent, on the steered answers. Two-way : answers steered toward aphasia, labeled as aphasic or not. Five-way : answers steered toward Broca’s or Wernicke’s aphasia, labeled with the classifier’s five classes. All : the answers of both sheets together, on which Cohen’s

\kappa

treats the labels of the two sheets as one set of seven. Consensus : the answers to which both experts give the same label. Comparisons with expert 2 leave out the 4 answers that expert 2 did not label. Bold marks the higher agreement of the classifier and the LLM judge.

		Classifier			LLM judge
Class	Experts	Calls	Prec.	Rec.	Calls	Prec.	Rec.
Control	$16$	$18$	$50$	$\boldsymbol{56}$	$14$	$\boldsymbol{57}$	$50$
Conduction or anomic	$22$	$28$	$\boldsymbol{36}$	$\boldsymbol{45}$	$12$	$17$	$9$
Broca	$70$	$68$	$\boldsymbol{74}$	$\boldsymbol{71}$	$40$	$57$	$33$
Wernicke	$64$	$36$	$\boldsymbol{75}$	$42$	$111$	$51$	$\boldsymbol{89}$
Model scrambles	$52$	$31$	$\boldsymbol{52}$	$\boldsymbol{31}$	$4$	$50$	$4$

Appendix

Table 28: Precision and recall of each class, in percent, for the classifier and the LLM judge on the 181 answers steered toward Broca’s or Wernicke’s aphasia. Experts : the answers that either expert puts in the class; an answer on which the experts disagree counts in both of their classes. Calls : the answers to which the classifier or the judge gives the class. Prec. : precision, the share of the calls that either expert shares. Rec. : recall, the share of the experts’ answers in the class that the classifier or the judge also puts there. Bold marks the higher value of the two.

Expert 2	Ctrl.	Cond.	Broca	Wern.	MS	Total
	Classifier
Control	$\boldsymbol{8}$	$3$	$0$	$0$	$2$	$13$
Conduction or anomic	$6$	$\boldsymbol{9}$	$0$	$3$	$1$	$19$
Broca	$0$	$1$	$\boldsymbol{49}$	$6$	$8$	$64$
Wernicke	$4$	$14$	$3$	$\boldsymbol{27}$	$11$	$59$
Model scrambles	$0$	$0$	$15$	$0$	$\boldsymbol{9}$	$24$

Appendix

Table 29: Labels of the classifier against those of expert 2 on the 179 answers steered toward Broca’s or Wernicke’s aphasia that expert 2 labeled. Rows give expert 2’s label and columns the classifier’s, and the bold diagonal counts the 102 answers on which the two agree. Ctrl. : control speech; Cond. : conduction or anomic aphasia; Wern. : Wernicke’s aphasia; MS : model scrambles.

	Aphasia		Broca		Wernicke		All three
Model	w/o MS	with MS	w/o MS	with MS	w/o MS	with MS	w/o MS	with MS
GPT-2 XL	$6/7$	$9/14$	$7/7$	$7/14$	$5/13$	$5/14$	$18/27$	$21/42$
Llama-2-7B-chat	$2/2$	$2/9$	$6/6$	$7/12$	$2/7$	$3/12$	$10/15$	$12/33$
Llama-2-13B-chat	$4/9$	$4/13$	$5/9$	$7/13$	$1/6$	$1/9$	$10/24$	$12/35$
Pythia-1B	$7/12$	$8/14$	$0/4$	$6/14$	$11/13$	$12/14$	$18/29$	$26/42$
Pythia-1.4B	$9/13$	$9/13$	$8/8$	$13/14$	$9/14$	$9/14$	$26/35$	$31/41$

Appendix

Table 30: Steering success by model: the number of steered answers that either expert labels with the steering target, out of the model’s labeled answers for that target, without and with the answers that the experts call model scrambles (MS). Each model has at most 14 answers per target, to 7 prompts with 2 steering vectors, one from the Cinderella and one from the sandwich transcripts, and loses the answers that are too short to score or that the LLM judge left without a reply. An answer that one expert labels with the target and the other calls a model scramble counts as a success with MS and is left out without MS, so a model can succeed more often with MS. Over all models, either expert labels the target for 59%, 68%, and 63% of the answers steered toward aphasia, Broca’s aphasia, and Wernicke’s aphasia without MS.

	Aphasia		Broca		Wernicke		All three
Prompt	w/o MS	with MS	w/o MS	with MS	w/o MS	with MS	w/o MS	with MS
Tell me as much of the story of Cinderella as you can.	$9/9$	$11/12$	$6/7$	$8/14$	$7/10$	$9/13$	$22/26$	$28/39$
Tell me how you would make a peanut butter and jelly sandwich.	$7/9$	$7/13$	$1/3$	$6/14$	$8/11$	$8/14$	$16/23$	$21/41$
Tell me about the best trip you ever took.	$3/11$	$4/14$	$7/11$	$8/13$	$5/9$	$7/12$	$15/31$	$19/39$
Describe a happy childhood memory.	$6/10$	$7/12$	$9/12$	$10/13$	$9/14$	$9/14$	$24/36$	$26/39$
Tell me about your first job.	$7/9$	$7/14$	$6/9$	$7/12$	$5/9$	$6/12$	$18/27$	$20/38$

Appendix

Table 31: Steering success by prompt: the number of steered answers that either expert labels with the steering target, out of the labeled answers to the prompt for that target, without and with the answers that the experts call model scrambles (MS). The first two prompts are those of the Cinderella and sandwich tasks, and the other five are the connected-text prompts of the Text Aphasia Battery. Each prompt has at most 14 answers per target, from 7 models with 2 steering vectors, one from the Cinderella and one from the sandwich transcripts. For the Cinderella and sandwich prompts, 7 of these 14 come from the vectors of the same task, whose answers chose the steering layer and strength. An answer that one expert labels with the target and the other calls a model scramble counts as a success with MS and is left out without MS.

Model and vector	Prompt	Classifier	LLM judge	Answer
Steered toward aphasia, labeled aphasic by both experts
GPT-2 XL, Cinderella vector	Cinderella †	Aphasic ( $0.87$ )	Aphasic	I don,t know that I like to tell it. It’s a good thing I didn. / I was in my. . .
Steered toward Broca’s aphasia, labeled Broca by both experts
Llama-2-13B-chat, Cinderella vector	Simple meal	Model scrambles ( $0.39$ )	Broca	Sure! Here are the steps to make a simple meal: / First, choose a recipe you like and ingredients list need . go shopping groc yes. Next, preheat oven or stove top cooking no. Then, chop vegetables car yes. After that, cook meat dish yes. Finally, serve hot enjoy eat.
Llama-2-7B-chat, Cinderella vector	First job	Broca ( $0.71$ )	Broca	sure! my job is good. okay.
Steered toward Wernicke’s aphasia, labeled Wernicke by both experts

Appendix

Table 32: Steered answers, verbatim, with the labels of the classifier and the LLM judge. In the first five, both experts label the answer with the steering target; in the last two, both experts call the answer a model scramble, while the classifier labels it Broca. Vector : the task whose transcripts gave the steering vector, and whose answers chose the steering layer and strength; † marks an answer to that task, the source task. Prompt : the Cinderella or sandwich task, or the topic of a connected-text prompt of the Text Aphasia Battery, “ Describe the steps to make a simple meal ” or “ Tell me about your first job ”. The classifier’s probability for its label follows the label; for the answer steered toward aphasia it is

P(\mathrm{aphasia})

. A slash marks one or more line breaks, and the bracketed text replaces 171 line breaks.

Figure 9: Mean target probability that steering back leaves in the 54 CLAMP lesions, for the lesion’s own direction and for three random directions. Removed : removal of the direction from the residual stream after every block.

r

: the strength of a push, which adds the direction after one of eight layers; own and random pushes add vectors of the same norm. Each mean pools the lesions, the layers of a push, and the random directions, over the answers long enough to score. Before steering back, the lesioned answers have a mean target probability of 0.711, and a lesion is undone when its answer no longer reaches the target: its target probability falls to 0.5 or below and the target is not its most probable class.

By model family
		Removal undoes			Probability left
			Random
	Sets	Own	$\geq 1$	All $3$	Own	Random
GPT-2	$5$	$5$	$5$	$4$	$0.000$	$0.052$
Llama-2-chat	$11$	$10$ †	$11$	$11$	$0.029$	$0.028$
Pythia	$22$	$22$	$22$	$16$	$0.005$	$0.133$

Appendix

Table 33: Removing a direction after every block of the 54 CLAMP lesions, by model family and by target. Removal undoes : the number of lesions whose answer no longer reaches the target, a target probability of 0.5 or below with the target not the most probable class, when the own direction is removed, and when at least one, or all three, of three random directions are removed. Probability left : the mean target probability after removal, against 0.711 before it. † With its own direction removed, the remaining lesion, of Llama-2-13B-chat on the sandwich task with the Wernicke target, gives an answer too short to score.

Intact	Lesioned	Lesioned and steered back
GPT-2 XL, Broca target, lesion of 146 heads; own direction pushed after layer 24 at $r=-0.5$
$P(\mathrm{Broca})=0.000$ , control	$P(\mathrm{Broca})=0.884$ , Broca	$P(\mathrm{Broca})=0.000$ , control
I’d start with the bread, cut it into slices about 1/2 inch thick, then spread some of your favorite jam on each slice. Then add in some sliced peanuts (or other nuts) and top that off with a scoop of PB&J. You can also use peanut butter instead of…	I’d love to have some peanut butter, but not the kind of peanut butter. / I’m sure do like peanut butter. / But I don’t know what it is. / It’s just peanut butter. / And I can’t tell you what it is. / Peanut Butter. / / That’s right? / Yes. / No. / What are you talking…	I’d use the same recipe as shown in this picture, but with only half of the peanut butter and half of the jelly. Then I’d put it on top side of the bread and then I’d spread it out to form a circle around the bread and I’d cut off about 1/4th of an…
Llama-2-7B-chat, aphasia target, lesion of 60 heads; own direction pushed after layer 8 at $r=-0.5$
$P(\mathrm{aphasia})=0.006$ , control	$P(\mathrm{aphasia})=0.931$ , Broca	$P(\mathrm{aphasia})=0.006$ , control
Sure! Here’s my step-by-step guide on how to make the perfect peanut butter and jelly sandwich: / / Ingredients: / / * 2 slices of bread (white or whole wheat, whichever you prefer) / * Peanut butter (creamy or crunchy, your choice) / * Jelly or jam (your…	[emoji] peanut Butt the jigglesand!	Sure! Here’s one way to make a peanut butter: / / / / 1. Start by spreading a layer of bread on the bottom of your pan. You can use any type of bread you like, such as white or whole wheat. Just be sure it is flat enough for the filling. / 2. Next, add a…

Appendix

Table 34: Two CLAMP lesions on the sandwich task that a weak push with the lesion’s own direction undoes. Each row gives the answers of the intact model, of the lesioned model, and of the lesioned model steered back by the push, each under its target probability and the classifier’s most probable class. We chose the two lesions because their answers read well, so they are not typical: in 16 of the 54 lesions, no push with the own direction undoes the lesion with a fluent answer that the classifier labels control speech, where a fluent answer does not loop and has at least 80% of its non-space characters in dictionary words. Answers are verbatim up to the last whole word within their first 250 characters; … marks the omitted rest, a slash marks a line break, and [emoji] replaces an emoji.

LLM Interpretability

Large language model (LLM) interpretability studies how internal representations and computations produce outputs, and how they can be explained or controlled.

See all 778 papers

May 15, 2026

Artificial Aphasias in Lesioned Language Models

Aphasias, selective language impairments which can arise from brain damage, reveal the functional organization of human language by providing causal links between affected brain regions and specific symptom profiles. Drawing on this literature, we introduce an aphasia-inspired technique to characterize the emergent functional organization of language models (LMs). We ``lesion'' (zero-out) model parameters and measure the effects of this intervention against clinical aphasia symptoms, as diagnosed by the Text Aphasia Battery (TAB). When applied to 112,426 outputs from five 1B-scale LMs, the full range of evaluated symptoms surface, but in distributions largely distinct from those of humans. Our method uncovers broad symptom-profile differences between attention components (query, key, value, output) and feed-forward components (up, gate, down), with weaker evidence for differences among components within the same mechanism. We also find an effect of depth, where lesions in early layers disproportionately cause syntactic and semantic symptoms while late-middle layers yield higher rates of phonological and fluency deficits. Although some LM lesions induce quantitatively more similar profiles to some human aphasia types than others, qualitative differences in symptom patterns between LMs and humans suggest that aphasia syndromes are heavily influenced by the details of learning and processing rather than being a domain-invariant consequence of disrupted language processing.

Nathan Roll, Jill Kries, Laura Gwilliams +1

Department of {1Linguistics, 2Psychology}, Stanford University · 3Wu Tsai Neurosciences Institute, Stanford University

LLM Interpretability Causal Interventions in Language Models

Ask PDF

Aug 13, 2026

Perturbation-based Regional Interpretability through Subtraction Mapping (PRISM): naming-error dissociations in language models and post-stroke aphasia

Mechanistic interpretability of large language models lacks spatially resolved, falsifiable tools for testing whether internal components are specialized for distinct cognitive operations. We adapt subtraction analysis, the standard framework of human neuroimaging, from biological brains to perturbed transformers, and apply the same logic to both substrates in parallel. Building on the Brain-LLM Unified Model (BLUM), which showed that layer-perturbed LLaVA-1.6-Vicuna-13B error profiles match the lesion patterns of aphasic patients, we develop PRISM (Perturbation-based Regional Interpretability through Subtraction Mapping). PRISM maps the seven clinical Philadelphia Naming Test categories, subtracts error classes pairwise, and treats each perturbation seed as a subject in a group analysis with threshold-free cluster enhancement along the layer axis. We run a structurally matched analysis on 213 chronic post-stroke aphasia patients using correlation-difference lesion-symptom mapping, and replicate both sides on held-out splits. The designs match in subject dimension (seeds, patients), spatial dimension (layers, atlas-parcellated cortex) and thresholding, but the contrast operator differs: a within-subject error-proportion difference for the LLM, a between-subject correlation difference for the cortex. Both substrates recover a robust phonemic-favoring dissociation, a deep layer cluster and a frontal-perisylvian cortical cluster, both replicating; the semantic-favoring direction is a consistently signed but non-significant trend on both. PRISM thus gives a falsifiable, spatially resolved test of functional-specialization claims in transformer language models. A confirmatory ROI-level intervention (PRISM Stage 3) licensing the strongest causal-mechanism claim is left to subsequent work.

Xiang Guan, Roger D. Newman-Norlund, Yong Yang +8

Department of Computer Science and Engineering, University of South Carolina · Department of Communication Sciences and Disorders, University of South Carolina · ALLT.AI, LLC, Columbia, SC +3

Transformer Interpretability LLM Interpretability Neural Network Interpretability

Ask PDF

Aug 5, 2026

Recovering Lesion Parameters from Aphasic Picture Naming Error Profiles in Large Language Models

Interpretability methods for large language models (LLMs) describe internal state but do not directly test whether that state is causally sufficient to produce the observed behavior. In earlier work, we lesioned LLMs to produce error profiles in picture naming, a central task for assessing aphasia, and found that specific lesions produced errors resembling those of individual stroke survivors. Here we ask the inverse question: given an error profile, can the lesion parameters that produced it be recovered, and what does this inverse problem reveal about transformer computation? Lesions in LLaVA-Vicuna 13B were parameterized by layer index, modification percentage, and noise sigma across 4,840 configurations, and error profiles were characterized by a seven-category clinical taxonomy (correct, semantic, unrelated, formal, mixed, neologism, no-response). We trained a multi-task neural network to map error profiles back to perturbation parameters. The problem admitted a partial solution: across 10 independently trained inverse models, modification percentage and noise sigma were recoverable, whereas layer index was recoverable only within a neighborhood. In counterfactual validation, a fresh model instance perturbed with the recovered parameters reproduced the target behavior in 81.4% of cases. This dissociation between low layer recovery and high counterfactual fidelity is consistent with functional redundancy across transformer layers, a property not captured by standard interpretability methods. As an out-of-distribution test, we applied the trained model to picture-naming error profiles from 278 stroke survivors; recovered parameters were syndrome-discriminative, most strongly for perturbation intensity, indicating generalization beyond the training distribution. Counterfactual validation provides a general framework for LLM interpretability claims beyond inverse mapping.

Yong Yang, Roger Newman-Norlund, Xiang Guan +10

Department of Computer Science and Engineering, University of South Carolina · Department of Communication Sciences and Disorders, University of South Carolina · ALLT.AI, LLC +3

LLM Interpretability Causal Interventions in Language Models Inverse Problems

Ask PDF

Abstract

LLM Interpretability

Explore similar work

Artificial Aphasias in Lesioned Language Models

Perturbation-based Regional Interpretability through Subtraction Mapping (PRISM): naming-error dissociations in language models and post-stroke aphasia

Recovering Lesion Parameters from Aphasic Picture Naming Error Profiles in Large Language Models