Can Language Models Get "Aphasia" From Neuronal Ablation?

Pratik Worah; Randall Bateman

Figure 1: Summary of ablation: clinical parallel mapping across the ablation experiments. Silencing Transformer feedforward (MLP/FFN) neurons, Transformer attention heads, and whole LSTM hidden units produces distinct failure modes. These patterns parallel focal aphasic syndromes versus steady diffuse decline.

Architecture	Ablated Computational Units	Params	Baseline Acc.
Small LSTM	Whole recurrent hidden units	${\sim}$ 10M	27.5%
AWD-LSTM	Whole recurrent hidden units	${\sim}$ 44M	38.1%
Qwen 0.5B	(1) Feedforward (MLP) neurons; (2) Attention heads	${\sim}$ 620M	53.6%
Qwen 1.8B	(1) Feedforward (MLP) neurons; (2) Attention heads	1.8B	54.7%

Table 1: Summary of model architectures and unit-level ablation targets. Baseline Acc. is the top-1 next-word prediction accuracy of the undamaged (

k=0

) model under the default protocol. Without the brief adaptation step, undamaged Qwen 0.5B and 1.8B score

51.5\%

and

50.6\%

, while the LSTMs are unchanged at

27.5\%

and

38.1\%

.

Figure 2: Next-word prediction accuracy relative to the undamaged baseline (

\mathrm{Acc}(k)/\mathrm{Acc}(0)

) across unit-ablation fractions

k\in[0,0.9]

(means

\pm 1

SE). The six conditions reveal three distinct degradation profiles across architectures and pathways. (a,b) Row 1: Whole-unit removal in both the Small LSTM and pretrained AWD-LSTM causes a steady, near-monotonic decline from

k=0.1

onward, with neither an early plateau nor a late collapse cliff. (c,e) Rows 2–3, left: Qwen 0.5B and 1.8B MLP neuron removal initially declines at a pace similar to the LSTMs through

k\leq 0.3

. It then undergoes a steep threshold collapse across

k^{*}\approx 0.3

–

0.4

to near-zero accuracy. (d,f) Rows 2–3, right: Qwen 0.5B and 1.8B attention-head removal shows early plateau-like buffering across

k=0.1

–

0.3

followed by a smoother S-shaped decline. Full numerical trajectories and curve-shape tests appear in Supplement S2 (Supplementary Tables 5 – 6 ), connection-pruning controls appear in Supplementary Figure 8 , and Supplementary Figure 9 plots the same trajectories against the fraction of non-embedding parameters removed.

Figure 3: Normalised moment ratio

R_{4}(k)/R_{4}(0)

(Equation 2 ) across unit-ablation fractions

k\in[0,0.9]

on a shared vertical scale (means

\pm 1

SE; token IDs are ordered by corpus frequency in the LSTMs and BPE merge order in Qwen). Because any drop in overall accuracy adds errors on common, low-index words (

i\approx 0

), raw

R_{4}(k)/R_{4}(0)

rises as accuracy falls, so each trajectory closely mirrors the inverse of its accuracy curve in Figure 2 (motivating the accuracy-matched null model in Figure 4 ). (a,b) Row 1: Whole-unit removal in both LSTMs produces a smooth, modest rise across the sweep (

\leq 5.1\%

and

\leq 7.7\%

). (c,e) Rows 2–3, left: Qwen MLP neuron removal changes little at mild damage (dipping slightly to

0.990

at

k=0.1

in Qwen 1.8B before rising to

1.031

–

1.037

at

k=0.3

), then jumps sharply across the

k^{*}\approx 0.3

–

0.4

collapse cliff. (d,f) Rows 2–3, right: Qwen attention-head removal stays nearly flat during the

k\in[0.1,0.3]

plateau (

\leq 1.9\%

change) before rising across

k=0.4

–

0.7

(more steeply in Qwen 0.5B than in Qwen 1.8B). Per-

k

values and connection-pruning controls appear in Supplement S3 and Supplementary Figure 10 .

Figure 4: Comparison of ablated error distributions against an accuracy-matched random-failure baseline (Equations 3 – 5 ). In all three panels, the horizontal axis is surviving relative accuracy

\mathrm{Acc}(k)/\mathrm{Acc}(0)

(from

1.0

at

k=0

on the left to

0.0

on the right), comparing models at equal accuracy loss. The vertical axes plot: (Left)

\Delta R_{4}=R_{4}^{\mathrm{obs}}-R_{4}^{\mathrm{null}}

, the difference in moment ratio relative to random failure; (Centre)

\Delta D_{1}=D_{1}^{\mathrm{obs}}-D_{1}^{\mathrm{null}}

, the difference in effective number of distinct error words relative to random failure; and (Right)

\mathrm{JS}(p(k)\,\|\,p(0))

, the Jensen–Shannon divergence between the damaged and undamaged (

k=0

) error distributions. Together, the three panels establish two main findings: first, neuronal ablation fails non-randomly across all models, concentrating fewer new mistakes on ultra-common words (

\Delta R_{4}<0

) and spreading errors across more distinct words (

\Delta D_{1}>0

) than random failure; second, Transformer ablation (especially MLP removal) reshapes the error distribution much more strongly than LSTM whole-unit removal (Supplement S3 ).

Figure 5: Syntactic and semantic performance on the 100-prompt fill-in-the-blank (Cloze) test under progressive unit-level ablation. Each panel plots grammatical part-of-speech match rate (blue circles, syntax) and GloVe cosine similarity to the target word (red squares, semantics) on the vertical axis against the fraction of units removed

k\in[0.0,0.9]

on the horizontal axis (means

\pm 1

SE across trials;

k=0.0

is the undamaged model). Row 1: Under whole-unit removal in both the Small LSTM (left) and pretrained AWD-LSTM (right), syntax and semantics erode in parallel from

k=0.1

onward, reaching near zero at

k=0.4

and

0.0\%

POS by

k=0.5

–

0.6

as outputs default to common function words. Rows 2–3: In both Qwen 0.5B and Qwen 1.8B, MLP neuron ablation (left) declines moderately through

k=0.3

before a steep threshold collapse across

k^{*}\approx 0.3

–

0.4

(near zero by

k=0.5

–

0.6

), whereas attention-head ablation (right) declines much more gradually on both metrics through

k=0.5

–

0.6

.

Architecture	Target	$k$	POS Match	Cos Sim	Trajectory	Qualitative Error Profile and Examples ( $k=0.2$ )
Baselines
Small LSTM	–	0.0	59.0%	0.337	–	grammatical continuations, no exact matches (4% correct, all via top-5, e.g. neck $\to$ head )
AWD-LSTM	–	0.0	57.0%	0.450	–	moderate accuracy (25% correct; time ✓, sky ✓, song ✓, ocean ✓)
Qwen 0.5B	–	0.0	65.0%	0.643	–	strong (64% correct; time ✓, dress ✓, hug ✓)
Qwen 1.8B	–	0.0	64.0%	0.663	–	strong (66% correct; time ✓, sky ✓, hug ✓)
Wernicke-type parallel: related-word substitutions at the intact rate, more no-responses & late threshold collapse

Table 2: Cloze syntactic (POS match) and semantic (GloVe cosine) benchmarks and representative errors across the six unit-level conditions, grouped by clinical parallel. Transformer MLP and attention pathways diverge sharply at severe damage (

k=0.5

), whereas both LSTMs erode syntax and semantics in lockstep until reaching the content-word floor across

k=0.4

–

0.6

(means across trials; full trajectories appear in Supplement S5 and Supplementary Table 17 ).

Architecture	Target	Quantitative Signature	Qualitative & Lexical Evidence	Clinical Parallel
Qwen 1.8B	MLP	Initial accuracy decline comparable to LSTMs ( $93.5\%$ and $82.0\%$ of baseline at $k{=}0.1$ , $0.2$ ), followed by late threshold collapse ( $k^{*}{\approx}0.3$ – $0.4$ ); largest negative shift in null-corrected $R_{4}$	Preserves the intact PNT Semantic rate ( $11.7\%$ vs. $11.0\%$ intact) while No-response rises ( $18.1\%$ vs. $1.0\%$ intact); produces within-category substitutions (“ocean” $\to$ “bay”, “sea” $\to$ “river”) and attribute substitutions (“cold” $\to$ “shallow”) with unchanged word concreteness	Wernicke-type
Qwen 0.5B	MLP	Same pattern: initial decline comparable to LSTMs ( $89.3\%$ and $70.6\%$ at $k{=}0.1$ , $0.2$ ) and late threshold collapse ( $k^{*}{\approx}0.3$ – $0.4$ ); strong negative shift in null-corrected $R_{4}$	Preserves the intact PNT Semantic rate ( $9.9\%$ vs. $9.0\%$ intact) while No-response rises ( $16.7\%$ vs. $1.0\%$ intact); produces contextually plausible substitutions (“cheer” $\to$ “sing”, “storm” $\to$ “dawn”) and generic evaluative defaults (“stunning” $\to$ “fine”)	Wernicke-type
Qwen 1.8B	Attn	Early retention at $k{=}0.1$ followed by a smoother sigmoid-like decline; preserves substantial Cloze syntax and semantics through $k{=}0.5$ ; milder null-corrected $R_{4}$ shift than MLP	Slightly fewer semantic paraphasias than intact ( $9.1\%$ vs. $11.0\%$ ) and a borderline shift toward more abstract words ( $\Delta\text{Conc.}{=}{+}0.24$ vs. ${+}0.13$ intact): generic substitutions (“gentle” $\to$ “small”) alongside within-domain ones (“keyboard” $\to$ “desktop”)	svPPA
Qwen 0.5B	Attn	Early retention at $k{=}0.1$ followed by a smoother sigmoid-like decline; Cloze syntax and semantics buffered through $k{=}0.4$ – $0.5$ ; milder null-corrected $R_{4}$ shift than MLP	Shifts substitutions toward more abstract words than the intact model ( $\Delta\text{Conc.}{=}{+}0.43$ vs. ${+}0.15$ intact), with generic defaults such as “ball” $\to$ “following” and “clear” $\to$ “good”	svPPA
AWD-LSTM	Whole units	Steady, gradual accuracy decline across $k\in[0,0.9]$ without a late cliff ( $75.3\%$ of baseline at $k{=}0.2$ , $34.4\%$ at $k{=}0.5$ ); modest gradual $R_{4}$ rise ( $\leq 7.7\%$ ; $\mathrm{JS}{=}0.008$ at $\mathrm{RelAcc}{=}0.80$ )	Syntax (POS) and semantics (cosine) erode in lockstep across $k{=}0.1$ – $0.3$ without dissociating ( $46.6\%$ content sub. at $k{=}0.2$ ), reaching the content-word Cloze floor across $k{=}0.4$ – $0.6$ ( $0.0\%$ POS by $k{=}0.6$ )	Diffuse AD-type
Small LSTM	Whole units	Steady, gradual accuracy decline across $k\in[0,0.9]$ without a late cliff ( $82.2\%$ of baseline at $k{=}0.2$ , $45.8\%$ at $k{=}0.5$ ); modest gradual $R_{4}$ rise ( $\leq 5.1\%$ ; $\mathrm{JS}{=}0.007$ at $\mathrm{RelAcc}{=}0.80$ )	Syntax and semantics erode in parallel across $k{=}0.1$ – $0.3$ without dissociating, progressing rapidly to non-content defaults ( $84.3\%$ non-content at $k{=}0.2$ ) and reaching the content-word Cloze floor across $k{=}0.4$ – $0.5$ ( $0.0\%$ POS by $k{=}0.5$ )	Diffuse AD-type

Table 3: Clinical parallel synthesis across the six unit-level ablation conditions. Each condition is classified along two converging axes: quantitative degradation dynamics (

k\in[0,0.9]

) and qualitative Cloze error profiles at

k=0.2

(detailed error-type breakdowns appear in Table 4 , Figure 6 , and Supplements S6 and S8 ).

Architecture	Condition	Correct	Content Substitution	Non-content	Mean $\Delta$ Conc.
Undamaged baselines ( $k=0$ , $n=100$ prompts)
Qwen 1.8B	Intact ( $k{=}0$ )	66.0%	27.0%	7.0%	$+0.13$
Qwen 0.5B	Intact ( $k{=}0$ )	64.0%	27.0%	9.0%	$+0.15$
AWD-LSTM	Intact ( $k{=}0$ )	25.0%	53.0%	22.0%	$+0.26$
Small LSTM	Intact ( $k{=}0$ )	4.0%	45.0%	51.0%	$+0.53$
Unit-ablated conditions at $k=0.2$ (mean [95% CI] across 10 independent ablation trials, $100$ prompts per trial)

Table 4: Cloze error-type distribution and Brysbaert concreteness drop (

\Delta\text{Conc.}

, target minus prediction ( Brysbaert et al., 2014 ) ; positive values indicate more abstract substitutions) for undamaged (

k=0

) and unit-ablated (

k=0.2

) models. MLP ablation lowers accuracy and raises non-content outputs while keeping substitution concreteness near the undamaged baseline, whereas attention ablation shifts substitutions toward more abstract words. For the ablated models, values show the mean and 95% confidence interval across 10 independent ablation trials (coding and exact-match details appear in Supplements S6 – S7 ).

Figure 6: Comparison of undamaged (

k=0

) and unit-ablated (

k=0.2

) model predictions with 278 post-stroke aphasia patients (PWA) ( Yang et al., 2026 ) on the Philadelphia Naming Test (PNT) taxonomy ( Roach et al., 1996 ; Mirman et al., 2010 ) (full seven-category numerical values appear in Supplementary Table 20 , Supplement S8 ). (a) Category proportions at

k=0.2

(mean

\pm

95% CI across 10 independent ablation trials,

100

prompts per trial) ordered by model scale against the post-stroke PWA reference, with the three sound-based categories (Formal, Mixed, and Neologism) grouped as Phonological. Qwen 1.8B MLP ablation preserves its undamaged rate of semantic paraphasias while No-response failures rise into the patient range, whereas LSTMs remain dominated by unrelated defaults. (b) Overall difference (Total Variation distance across all seven PNT categories, in percentage points) between the ablated (

k=0.2

) unit/MLP model predictions and the post-stroke patient percentages as model size increases.

Figure 7: Schematic overview of model architectures and ablation modes. (a) In the 3-layer Small LSTM and pretrained AWD-LSTM (serial layer stack without residual skip connections), our primary target is whole-unit removal (left card), which zeroes weight_ih rows, weight_hh incoming rows and outgoing columns, gate biases bias_ih and bias_hh , and outgoing columns in the next layer’s weight_ih to silence unit

j

completely (

h_{t,j}\equiv 0

). To separate the two incoming pathways (Section 3 ; Supplements S2 – S6 ), we also evaluate two connection-level controls: recurrent-connection pruning (centre card), which zeroes weight_hh rows/columns and gate biases while unit

j

still receives current input

x_{t}

, and input-connection pruning (right card), which zeroes weight_ih rows and gate biases while unit

j

still receives recurrent state

h_{t-1}

. (b) In the 24-layer Qwen 0.5B and Qwen 1.8B Transformers, a residual stream (

+

) bypasses each layer’s attention and MLP sub-blocks. We compare (1) attention-head removal (

Q,K,V

projection rows/biases and

O

projection columns zeroed) and (2) MLP neuron removal ( up_proj / gate_proj rows and down_proj columns zeroed).

Model	Ablation	Acc@ $k{=}0$	Acc@ $k{=}0.5$	RelAcc	Clinical Parallel
Small LSTM	Unit (w_ih + w_hh)	0.2747	0.1260	0.458	Diffuse AD-type ( $\to$ global collapse)
AWD-LSTM	Unit (w_ih + w_hh)	0.3806	0.1307	0.344	Diffuse AD-type ( $\to$ global collapse)
Qwen 0.5B	MLP neurons	0.5363	0.0417	0.078	Wernicke-type
Qwen 0.5B	Attention heads	0.5363	0.1067	0.199	svPPA
Qwen 1.8B	MLP neurons	0.5465	0.1163	0.213	Wernicke-type
Qwen 1.8B	Attention heads	0.5465	0.1732	0.317	svPPA

Table 5: Full ablation summary at

k=0.5

across the six primary unit-level conditions and the four supplementary LSTM connection-pruning controls.

Model	Lesion	Axis $x$	Linear Slope	$\Delta\mathrm{AIC}$	Fit Verdict (Qualitative Shape)
Small LSTM	Unit removal (w_ih + w_hh)	$k$	$-0.888$	$-39.04$	non-linear (gradual bend)
Small LSTM	Unit removal (w_ih + w_hh)	frac. params	$-0.812$	$-23.71$	non-linear (gradual bend)
Small LSTM	Recurrent-connection (w_hh)	$k$	$-0.556$	$-26.23$	non-linear (gradual bend)
Small LSTM	Recurrent-connection (w_hh)	frac. params	$-1.019$	$-7.75$	non-linear (nearly straight line)
Small LSTM	Input-connection (w_ih)	$k$	$-0.793$	$-104.98$	non-linear (steep early drop)
Small LSTM	Input-connection (w_ih)	frac. params	$-1.577$	$-106.07$	non-linear (steep early drop)

Table 6: Curve-shape comparison across trial-level relative accuracy points. Each row compares a 3-parameter S-curve (logistic) fit

y(x)=a/(1+\exp(b(x-x_{0})))

against a 2-parameter straight-line fit

y(x)=a+bx

as a function of

k

or fraction of non-embedding parameters removed. Because the 3-parameter curve bends toward zero at high

k

, any trajectory that curves (whether through an initial flat plateau followed by a cliff, a steep early drop, or a smooth bend) achieves

\Delta\mathrm{AIC}=\mathrm{AIC}_{\mathrm{logistic}}-\mathrm{AIC}_{\mathrm{linear}}<-2

(labelled non-linear in the table). The rightmost column notes the qualitative shape seen in Figures 2 and 9 .

Figure 8: Relative accuracy (

\mathrm{Acc}(k)/\mathrm{Acc}(0)

, normalised to baseline) degradation curves for the supplementary LSTM connection-pruning controls (top two rows: Small LSTM and AWD-LSTM recurrent-connection pruning weight_hh in the left column and input-connection pruning weight_ih in the right column) alongside the Qwen 0.5B and Qwen 1.8B MLP and Attention pathways (bottom two rows; primary LSTM whole-unit curves are shown in Figure 2 a,b). Isolating weight_hh vs. weight_ih shows that recurrent-connection pruning degrades gradually through

k=0.9

, whereas input-connection pruning collapses steeply by

k=0.2

–

0.4

. All curves show means

\pm 1

SE across trials.

Figure 9: Relative accuracy plotted against nominal unit/connection fraction

k

(left column) and against the exact fraction of non-embedding parameters removed (right column) across all four architectures, comparing whole-unit removal ( weight_ih + weight_hh , solid brown curves) against recurrent-connection pruning ( weight_hh , coral in row 1 and dark green in row 2) and input-connection pruning ( weight_ih , medium sea green in row 1 and teal in row 2) in both LSTMs, and MLP vs. attention-head removal in both Qwen models (rows 3–4). Plotted against parameter fraction, attention-head ablation degrades faster per parameter removed than MLP ablation in both Qwen models, and both Qwen pathways degrade faster per parameter removed than LSTM whole-unit removal and recurrent-connection pruning, which decline monotonically without a cliff.

Figure 10: Moment ratio

R_{4}(k)/R_{4}(0)

for the supplementary LSTM connection-pruning pathways (top two rows: Small LSTM and AWD-LSTM recurrent-connection pruning weight_hh in the left column and input-connection pruning weight_ih in the right column) alongside Qwen 0.5B and Qwen 1.8B MLP and Attention pathways (bottom two rows; shared vertical scale; means

\pm 1

SE; primary LSTM whole-unit

R_{4}

curves are shown in Figure 3 a,b).

Figure 11: Fourier spectral analysis of the frequency-ranked error distribution across unit-ablation fractions

k\in[0,0.9]

for the six primary unit-level conditions (means

\pm 1

SD): within each panel, the left graph shows spectral slope

\beta

(orange) and the right graph shows the low-to-high-frequency spectral gap

G

(teal). Row 1: Small LSTM whole-unit removal (left) and AWD-LSTM whole-unit removal (right). Rows 2–3: Qwen 0.5B and Qwen 1.8B MLP neuron ablation (left) and attention-head ablation (right), where Qwen 1.8B and Qwen 0.5B attention remain buffered through

k\leq 0.3

–

0.4

before shifting at higher

k

, while Qwen 0.5B MLP and AWD-LSTM whole-unit removal show non-monotonic fluctuations across

k

(per-

k

values in Tables 7 – 16 ).

Figure 12: Spectral slope

\beta

(orange, left graph of each pair) and spectral gap

G

(teal, right graph of each pair) across

k\in[0,0.9]

for the supplementary LSTM connection-pruning controls (top two rows: Small LSTM and AWD-LSTM recurrent-connection pruning weight_hh in the left column and input-connection pruning weight_ih in the right column) alongside Qwen 0.5B and Qwen 1.8B MLP and Attention pathways (bottom two rows; means

\pm 1

SD).

$k$	Accuracy	$\pm$ SE	RelAcc	$R_{4}$	$\pm R_{4}$ SE	$R_{4}/R_{4}(0)$	$\beta$	$\pm\beta$ SD	$G$
0.00	0.3806	0.0000	1.0000	2.2084	0.0000	1.0000	$-$ 1.5328	0.0002	3.3106
0.10	0.3410	0.0007	0.8960	2.2153	0.0014	1.0031	$-$ 1.3621	0.1268	3.2343
0.20	0.2866	0.0009	0.7531	2.2355	0.0009	1.0123	$-$ 1.3563	0.1657	3.5633
0.30	0.2295	0.0041	0.6030	2.2684	0.0028	1.0272	$-$ 1.2366	0.1620	3.5111
0.40	0.1731	0.0046	0.4548	2.3059	0.0031	1.0441	$-$ 1.1410	0.1227	3.4511
0.50	0.1307	0.0049	0.3435	2.3345	0.0033	1.0571	$-$ 1.4492	0.2976	4.3659

Table 7: AWD-LSTM Whole-Unit Removal ( weight_ih + weight_hh + biases): full metrics.

$k$	Accuracy	$\pm$ SE	RelAcc	$R_{4}$	$\pm R_{4}$ SE	$R_{4}/R_{4}(0)$	$\beta$	$\pm\beta$ SD	$G$
0.00	0.2747	0.0000	1.0000	2.2575	0.0000	1.0000	$-$ 1.3724	0.0001	3.7113
0.10	0.2548	0.0008	0.9274	2.2661	0.0005	1.0038	$-$ 1.3352	0.0652	3.6480
0.20	0.2257	0.0015	0.8215	2.2810	0.0017	1.0104	$-$ 1.3652	0.1346	3.9358
0.30	0.1930	0.0011	0.7024	2.2955	0.0014	1.0168	$-$ 1.3473	0.2175	3.8458
0.40	0.1565	0.0052	0.5695	2.3177	0.0035	1.0267	$-$ 1.3513	0.3677	3.9571
0.50	0.1260	0.0057	0.4585	2.3377	0.0038	1.0355	$-$ 1.2225	0.5036	3.7887

Table 8: Small LSTM Whole-Unit Removal ( weight_ih + weight_hh + biases): full metrics. At

k=0.90

, all five trials settle on predicting the single most common word in the test text ( “the” ,

7.12\%

of tokens), giving zero variation across trials (

\pm 0.0000

).

$k$	Accuracy	$\pm$ SE	RelAcc	$R_{4}$	$\pm R_{4}$ SE	$R_{4}/R_{4}(0)$	$\beta$	$\pm\beta$ SD	$G$
0.00	0.3806	0.0000	1.0000	2.2084	0.0000	1.0000	$-$ 1.5327	0.0002	3.3103
0.10	0.3595	0.0005	0.9445	2.2148	0.0005	1.0029	$-$ 1.4866	0.0375	3.2828
0.20	0.3303	0.0018	0.8678	2.2260	0.0014	1.0080	$-$ 1.3900	0.1209	3.1562
0.30	0.2941	0.0018	0.7727	2.2431	0.0014	1.0157	$-$ 1.4368	0.0510	3.5124
0.40	0.2611	0.0022	0.6861	2.2591	0.0019	1.0229	$-$ 1.5096	0.0874	3.7748
0.50	0.2193	0.0010	0.5763	2.2848	0.0012	1.0346	$-$ 1.5491	0.0695	4.2473

Table 9: AWD-LSTM Hidden ( weight_hh ) connection pruning (supplementary control 1): full metrics.

$k$	Accuracy	$\pm$ SE	RelAcc	$R_{4}$	$\pm R_{4}$ SE	$R_{4}/R_{4}(0)$	$\beta$	$\pm\beta$ SD	$G$
0.00	0.3806	0.0000	1.0000	2.2085	0.0000	1.0000	$-$ 1.5329	0.0001	3.3108
0.10	0.3395	0.0020	0.8921	2.2182	0.0023	1.0044	$-$ 1.4784	0.1623	3.5164
0.20	0.2768	0.0047	0.7274	2.2448	0.0030	1.0164	$-$ 1.3600	0.1614	3.5278
0.30	0.1847	0.0045	0.4852	2.2993	0.0031	1.0411	$-$ 1.4685	0.1298	4.3924
0.40	0.0966	0.0051	0.2538	2.3572	0.0033	1.0673	$-$ 1.6801	0.2852	5.0464
0.50	0.0570	0.0049	0.1498	2.3825	0.0031	1.0788	$-$ 1.5833	0.3095	4.7614

Table 10: AWD-LSTM Input ( weight_ih ) connection pruning (supplementary control 2): full metrics.

$k$	Accuracy	$\pm$ SE	RelAcc	$R_{4}$	$\pm R_{4}$ SE	$R_{4}/R_{4}(0)$	$\beta$	$\pm\beta$ SD	$G$
0.00	0.2747	0.0000	1.0000	2.2575	0.0000	1.0000	$-$ 1.3724	0.0001	3.7112
0.10	0.2517	0.0025	0.9161	2.2695	0.0015	1.0053	$-$ 1.3314	0.0194	3.5717
0.20	0.2131	0.0043	0.7755	2.2937	0.0033	1.0160	$-$ 1.2073	0.1430	3.3351
0.30	0.1908	0.0032	0.6945	2.3040	0.0027	1.0206	$-$ 1.0916	0.2242	2.9486
0.40	0.1819	0.0045	0.6621	2.3037	0.0026	1.0205	$-$ 1.4068	0.1665	3.8097
0.50	0.1593	0.0023	0.5798	2.3168	0.0015	1.0263	$-$ 1.4565	0.2034	3.8150

Table 11: Small LSTM Recurrent-Connection Pruning ( weight_hh , supplementary control 1): full metrics.

$k$	Accuracy	$\pm$ SE	RelAcc	$R_{4}$	$\pm R_{4}$ SE	$R_{4}/R_{4}(0)$	$\beta$	$\pm\beta$ SD	$G$
0.00	0.2747	0.0000	1.0000	2.2575	0.0000	1.0000	$-$ 1.3724	0.0001	3.7115
0.10	0.1786	0.0065	0.6500	2.3078	0.0046	1.0223	$-$ 1.2667	0.1908	3.4870
0.20	0.0380	0.0046	0.1385	2.3945	0.0028	1.0607	$-$ 1.2772	0.0980	3.4318
0.30	0.0333	0.0065	0.1214	2.3973	0.0040	1.0619	$-$ 1.1726	0.1351	3.2326
0.40	0.0266	0.0047	0.0968	2.4016	0.0029	1.0638	$-$ 1.1669	0.0633	3.1909
0.50	0.0157	0.0036	0.0572	2.4082	0.0022	1.0668	$-$ 1.1419	0.0177	3.1727

Table 12: Small LSTM Input-Connection Pruning ( weight_ih , supplementary control 2): full metrics.

$k$	Accuracy	$\pm$ SE	RelAcc	$R_{4}$	$\pm R_{4}$ SE	$R_{4}/R_{4}(0)$	$\beta$	$\pm\beta$ SD	$G$
0.00	0.5363	0.0000	1.0000	1.8861	0.0000	1.0000	$-$ 0.6608	0.0000	1.6345
0.10	0.4790	0.0060	0.8932	1.8908	0.0069	1.0025	$-$ 0.7806	0.0481	2.0512
0.20	0.3786	0.0117	0.7059	1.9155	0.0093	1.0156	$-$ 0.6893	0.0336	1.7207
0.30	0.3035	0.0089	0.5658	1.9558	0.0070	1.0370	$-$ 0.7995	0.1359	2.0021
0.40	0.1382	0.0221	0.2576	2.0416	0.0131	1.0825	$-$ 0.8825	0.1356	2.3562
0.50	0.0417	0.0103	0.0777	2.0957	0.0053	1.1111	$-$ 0.8690	0.0335	2.3038

Table 13: Qwen 0.5B MLP Neuron Removal: full metrics.

$k$	Accuracy	$\pm$ SE	RelAcc	$R_{4}$	$\pm R_{4}$ SE	$R_{4}/R_{4}(0)$	$\beta$	$\pm\beta$ SD	$G$
0.00	0.5363	0.0000	1.0000	1.8861	0.0000	1.0000	$-$ 0.6608	0.0000	1.6345
0.10	0.5232	0.0011	0.9756	1.8923	0.0027	1.0033	$-$ 0.6503	0.0196	1.6131
0.20	0.4536	0.0135	0.8458	1.9225	0.0102	1.0193	$-$ 0.6334	0.0682	1.5857
0.30	0.4320	0.0236	0.8055	1.9079	0.0165	1.0116	$-$ 0.6510	0.0258	1.6377
0.40	0.2608	0.0438	0.4863	1.9880	0.0245	1.0541	$-$ 0.6520	0.0875	1.4967
0.50	0.1067	0.0395	0.1990	2.0635	0.0203	1.0941	$-$ 0.7480	0.0646	1.7933

Table 14: Qwen 0.5B Attention-Head Removal: full metrics.

$k$	Accuracy	$\pm$ SE	RelAcc	$R_{4}$	$\pm R_{4}$ SE	$R_{4}/R_{4}(0)$	$\beta$	$\pm\beta$ SD	$G$
0.00	0.5465	0.0000	1.0000	1.8874	0.0000	1.0000	$-$ 0.6225	0.0000	1.5568
0.10	0.5112	0.0022	0.9354	1.8691	0.0031	0.9903	$-$ 0.6697	0.0334	1.7113
0.20	0.4479	0.0054	0.8197	1.8978	0.0065	1.0055	$-$ 0.6710	0.0687	1.5578
0.30	0.3376	0.0294	0.6178	1.9462	0.0164	1.0312	$-$ 0.7183	0.0708	1.6114
0.40	0.1287	0.0360	0.2355	2.0513	0.0184	1.0868	$-$ 0.8472	0.1312	2.2284
0.50	0.1163	0.0205	0.2129	2.0556	0.0108	1.0891	$-$ 0.9153	0.1664	2.3113

Table 15: Qwen 1.8B MLP Neuron Removal: full metrics.

$k$	Accuracy	$\pm$ SE	RelAcc	$R_{4}$	$\pm R_{4}$ SE	$R_{4}/R_{4}(0)$	$\beta$	$\pm\beta$ SD	$G$
0.00	0.5465	0.0000	1.0000	1.8874	0.0000	1.0000	$-$ 0.6225	0.0000	1.5568
0.10	0.5227	0.0059	0.9566	1.8823	0.0128	0.9973	$-$ 0.6022	0.0310	1.5457
0.20	0.4507	0.0432	0.8248	1.9104	0.0244	1.0122	$-$ 0.6507	0.0699	1.6066
0.30	0.4440	0.0171	0.8125	1.9127	0.0085	1.0134	$-$ 0.6640	0.0558	1.6601
0.40	0.3559	0.0196	0.6514	1.9593	0.0167	1.0381	$-$ 0.5792	0.0615	1.3152
0.50	0.1732	0.0376	0.3169	2.0397	0.0193	1.0807	$-$ 0.7272	0.1497	1.7441

Table 16: Qwen 1.8B Attention-Head Removal: full metrics.

Figure 13: Syntactic (POS match, blue) and semantic (GloVe cosine similarity, red) Cloze performance for the supplementary LSTM connection-pruning controls (rows 1–2: Small LSTM and AWD-LSTM recurrent-connection pruning weight_hh in the left column and input-connection pruning weight_ih in the right column) alongside Qwen 0.5B and Qwen 1.8B MLP and Attention pathways (rows 3–4; primary unit-level Cloze trajectories are shown in Figure 5 ). Isolating weight_hh (rows 1–2, left) confirms that pruning recurrent memory connections alone produces a gradual, non-dissociated decline all the way to

k=0.9

, whereas pruning weight_ih input connections (rows 1–2, right) drives the rapid bilateral Cloze collapse seen at high

k

under whole-unit removal.

		POS Match (Syntax)				Cosine Similarity (Semantics)
Model	Target	$k{=}0$	$k{=}0.2$	$k{=}0.5$	$k{=}0.9$	$k{=}0$	$k{=}0.2$	$k{=}0.5$	$k{=}0.9$
Small LSTM	unit	0.590	0.236	0.000	0.000	0.337	0.177	0.000	0.000
AWD-LSTM	unit	0.570	0.427	0.022	0.000	0.450	0.322	0.020	0.000
Qwen 0.5B	MLP	0.650	0.472	0.064	0.084	0.643	0.421	0.042	0.031
Qwen 0.5B	Attn	0.650	0.522	0.350	0.192	0.643	0.519	0.198	0.008
Qwen 1.8B	MLP	0.640	0.488	0.062	0.058	0.663	0.449	0.048	0.042

Table 17: Quantified Cloze test: POS match rate and GloVe cosine similarity at selected ablation fractions across all ten conditions (six primary unit-level conditions plus four supplementary LSTM connection-pruning controls). Values are means across trials (deterministic baseline at

k=0.0

,

n=10

trials at

k=0.2

, and

n=5

trials at

k\in\{0.5,0.9\}

).

Prompt	Exp.	Qwen1.8 Intact	Qwen1.8 MLP	Qwen1.8 Attn	AWD Unit	LSTM Unit	AWD Hid.	LSTM Hid.
once upon a	time	time $\checkmark$	day $\sim$	time $\checkmark$	a $\times$	” $\times$	time $\checkmark$	new $\times$
the sun was shining bright in the	sky	sky $\checkmark$	sky $\checkmark$	sky $\checkmark$	summer $\times$	first $\times$	sky $\checkmark$	same $\times$
she picked up the	ball	phone $\sim$	ball $\checkmark$	1 $\times$	number $\times$	first $\times$	first $\times$	end $\times$
the bird flew up into the	sky	sky $\checkmark$	trees $\sim$	tree $\times$	air $\times$	first $\times$	air $\sim$	first $\times$
she sang a beautiful	song	song $\checkmark$	song $\checkmark$	song $\checkmark$	, $\sim$	, $\times$	song $\checkmark$	- $\times$
he wanted to go	home	to $\sim$	to $\times$	to $\times$	on $\times$	to $\times$	to $\times$	to $\times$

Table 18: Representative qualitative Cloze predictions across architectures at

k=0.2

(from the first ablation trial), showing the undamaged (

k=0

) Qwen 1.8B prediction ( Qwen1.8 Intact ) alongside Qwen 1.8B MLP and Attention, primary LSTM whole-unit removal ( AWD Unit , LSTM Unit ), and supplementary LSTM recurrent-connection pruning ( AWD Hid. , LSTM Hid. ). Errors quoted as examples in the main text were checked against the top-1 and top-5 predictions of the corresponding undamaged model for the same prompt and do not occur there; some ablated errors in this table (for example flower

\to

girl and mouse

\to

number ) are also produced by the undamaged model and should not be read as ablation-induced.

\checkmark

= exact match;

\sim

= in top-5;

\times

= miss.

		Three-Way Cloze Error & Concreteness Profile				Seven-Category PNT Taxonomy (%)
Architecture	Pathway ( $k{=}0.2$ )	Correct	Content Sub.	Non-content	Mean $\Delta$ Conc.	Corr.	Sem.	Form.	Mix.	Unrel.	Neol.	No-resp.
AWD-LSTM	w_hh (recurrent)	20.2% [18.7, 21.7]	53.1% [51.0, 55.2]	26.7% [24.7, 28.7]	$+0.34$ [ $+0.27$ , $+0.41$ ]	20.2%	21.6%	0.1%	0.0%	54.9%	0.0%	3.2%
Small LSTM	w_hh (recurrent)	1.8% [1.1, 2.5]	37.2% [32.5, 41.9]	61.0% [55.9, 66.1]	$+0.67$ [ $+0.54$ , $+0.80$ ]	1.8%	7.8%	0.9%	0.0%	81.5%	0.0%	8.0%
AWD-LSTM	w_ih (input)	10.7% [8.8, 12.6]	53.9% [47.4, 60.4]	35.4% [28.5, 42.3]	$+0.30$ [ $+0.18$ , $+0.42$ ]	10.7%	18.4%	0.2%	0.0%	64.7%	0.0%	6.0%
Small LSTM	w_ih (input)	0.7% [0.4, 1.1]	23.0% [11.6, 34.4]	76.3% [65.0, 87.6]	$+0.85$ [ $+0.32$ , $+1.39$ ]	0.7%	1.6%	0.1%	0.0%	42.1%	0.0%	55.5%

Table 19: Error-type distribution, Brysbaert concreteness shift

\Delta\text{Conc.}

, and seven-category Philadelphia Naming Test (PNT) taxonomy at

k=0.2

for the four supplementary LSTM connection-pruning controls ( weight_hh recurrent pruning and weight_ih input pruning; compare with the primary unit-level results in Table 4 , Figure 6 , and Supplementary Table 20 ). Values show the mean and 95% confidence interval (or mean percentage for the PNT categories) across 10 independent ablation trials (

100

prompts per trial).

Undamaged baselines ( $k=0$ , $n=100$ prompts)
Architecture	Condition	Correct	Semantic	Formal	Mixed	Unrelated	Neologism	No-response
Qwen 1.8B	Intact ( $k{=}0$ )	66.0%	11.0%	0.0%	0.0%	22.0%	0.0%	1.0%
Qwen 0.5B	Intact ( $k{=}0$ )	64.0%	9.0%	0.0%	0.0%	26.0%	0.0%	1.0%
AWD-LSTM	Intact ( $k{=}0$ )	25.0%	19.0%	0.0%	0.0%	53.0%	0.0%	3.0%
Small LSTM	Intact ( $k{=}0$ )	4.0%	11.0%	1.0%	0.0%	80.0%	0.0%	4.0%
Unit-ablated conditions at $k=0.2$ (mean across 10 independent ablation trials, $100$ prompts per trial)

Table 20: Comparison of undamaged (

k=0

) and unit-ablated (

k=0.2

) model predictions with 278 post-stroke aphasia patients (PWA) [ Yang et al., 2026 ] on the seven-category Philadelphia Naming Test (PNT) taxonomy [ Roach et al., 1996 , Mirman et al., 2010 ] (visualised in Figure 6 in Section 2.5 ). At

k=0.2

, Qwen 1.8B MLP ablation preserves its undamaged rate of semantic paraphasias while non-word and punctuation failures (No-response) rise into the patient range, whereas LSTMs collapse onto unrelated defaults. For the ablated models, values are mean percentages across 10 independent ablation trials (

100

prompts per trial; 95% confidence intervals appear in Supplement S6 ).

Architecture	Pathway	0.50	0.55	0.60	0.65	0.70	0.75	0.80
Qwen 1.8B	MLP	27	18	18	14	10	7	6
Qwen 1.8B	Attn	21	12	11	9	7	4	3
Qwen 0.5B	MLP	17	11	10	8	6	5	3
Qwen 0.5B	Attn	13	7	5	4	2	2	1

Table 21: Number of predictions classified as Semantic (out of 100 in the first ablation trial) as a function of the semantic threshold

\theta_{s}

, with

\theta_{p}

fixed at 0.50.

Figure 14: Fine-tuning sensitivity across all three ablation modes of the Small LSTM (

k\in[0,0.9]

across all four arms): (a) recurrent-connection pruning ( weight_hh ), (b) input-connection pruning ( weight_ih ), and (c) whole-unit removal ( weight_ih + weight_hh , primary unit-level target). Curves compare zero fine-tuning ( none ), brief local SGD ( weak , default protocol), 500-step AdamW on TinyStories ( strong ), and 500-step AdamW on in-domain WikiText-103 ( strong_wt103 ). As in the AWD-LSTM, none and weak are nearly indistinguishable, and whole-unit and recurrent ablation decline smoothly without a buffered plateau across all four arms.

Figure 15: Fine-tuning sensitivity across all three ablation modes of the pretrained AWD-LSTM: (a) recurrent-connection pruning ( weight_hh ), (b) input-connection pruning ( weight_ih ), and (c) whole-unit removal ( weight_ih + weight_hh , primary unit-level target). Curves compare zero fine-tuning ( none ), brief local SGD ( weak , default protocol), 500-step AdamW on TinyStories ( strong ), and 500-step AdamW on in-domain WikiText-103 ( strong_wt103 ). Across all arms, whole-unit and recurrent ablation decline smoothly and monotonically without a buffered plateau.

Figure 16: Fine-tuning sensitivity across both pathways of the pretrained Qwen 0.5B Transformer (

k\in[0,0.9]

across all three arms): (a) MLP neuron removal and (b) attention-head removal. Under zero fine-tuning ( none ) and brief local SGD ( weak , 62 updates), MLP ablation declines at mild damage and then collapses sharply at

k\geq 0.4

, while attention ablation shows an early plateau before declining, whereas 500-step global AdamW ( strong ) re-optimises surviving residual pathways across layers to restore near-baseline accuracy.

Prompt	Exp.	Pred.		Top-5
once upon a	time	time	$\checkmark$	[time, midnight, lonely, summer, winter]
the sun was shining bright in the	sky	sky	$\checkmark$	[sky, clear, blue, garden, morning]
she picked up the	ball	phone	$\sim$	[phone, book, pen, ball, bag]
he put the food in his	mouth	mouth	$\checkmark$	[mouth, bag, pocket, hand, basket]
the little girl was wearing a pretty	dress	dress	$\checkmark$	[dress, skirt, white, pink, red]
the bird flew up into the	sky	sky	$\checkmark$	[sky, air, tree, trees, blue]

Table 22: Qwen 1.8B: Baseline (

k=0

), all 100 Cloze prompts.

Prompt	Exp.	Pred.		Top-5
once upon a	time	day	$\sim$	[day, summer, time, hill, night]
the sun was shining bright in the	sky	sky	$\checkmark$	[sky, clouds, , day, blue]
she picked up the	ball	ball	$\checkmark$	[ball, book, , pen, phone]
he put the food in his	mouth	mouth	$\checkmark$	[mouth, stomach, , belly, own]
the little girl was wearing a pretty	dress	dress	$\checkmark$	[dress, , (, red, ______]
the bird flew up into the	sky	trees	$\sim$	[trees, sky, air, tree, tall]

Table 23: Qwen 1.8B: MLP Ablation (

k=0.2

), all 100 Cloze prompts.

Prompt	Exp.	Pred.		Top-5
once upon a	time	time	$\checkmark$	[time, child, journey, traveler, great]
the sun was shining bright in the	sky	sky	$\checkmark$	[sky, park, morning, room, picture]
she picked up the	ball	1	$\times$	[1, 0, 2, 3, ]
he put the food in his	mouth	mouth	$\checkmark$	[mouth, hands, hand, stomach, head]
the little girl was wearing a pretty	dress	dress	$\checkmark$	[dress, pink, blue, shirt, white]
the bird flew up into the	sky	tree	$\times$	[tree, air, window, house, branches]

Table 24: Qwen 1.8B: Attention Ablation (

k=0.2

), all 100 Cloze prompts.

Prompt	Exp.	Pred.		Top-5
once upon a	time	time	$\checkmark$	[time, midnight, river, night, mountain]
the sun was shining bright in the	sky	sky	$\checkmark$	[sky, morning, east, clear, west]
she picked up the	ball	book	$\times$	[book, phone, key, bag, keys]
he put the food in his	mouth	mouth	$\checkmark$	[mouth, bag, pocket, stomach, hand]
the little girl was wearing a pretty	dress	dress	$\checkmark$	[dress, pink, red, blue, hat]
the bird flew up into the	sky	sky	$\checkmark$	[sky, air, tree, trees, blue]

Table 25: Qwen 0.5B: Baseline (

k=0

), all 100 Cloze prompts.

Prompt	Exp.	Pred.		Top-5
once upon a	time	once	$\times$	[once, before, in, when, a]
the sun was shining bright in the	sky	sky	$\checkmark$	[sky, air, sun, room, morning]
she picked up the	ball	things	$\times$	[things, new, work, book, song]
he put the food in his	mouth	’s	$\sim$	[’s, room, bag, mouth, ’s]
the little girl was wearing a pretty	dress	girl	$\times$	[girl, little, boy, and, .]
the bird flew up into the	sky	sky	$\checkmark$	[sky, air, room, clouds, park]

Table 26: Qwen 0.5B: MLP Ablation (

k=0.2

), all 100 Cloze prompts.

Prompt	Exp.	Pred.		Top-5
once upon a	time	time	$\checkmark$	[time, man, young, girl, woman]
the sun was shining bright in the	sky		$\times$	[, <?> , -, [CJK] , [CJK][CJK] ]
she picked up the	ball	following	$\times$	[following, words, right, photos, food]
he put the food in his	mouth	oven	$\sim$	[oven, car, room, bag, mouth]
the little girl was wearing a pretty	dress	much	$\times$	[much, , , a, ,]
the bird flew up into the	sky		$\times$	[, <?> , [CJK] , -, ]

Table 27: Qwen 0.5B: Attention Ablation (

k=0.2

), all 100 Cloze prompts.

Prompt	Exp.	Pred.		Top-5
once upon a	time	time	$\checkmark$	[time, television, live, single, young]
the sun was shining bright in the	sky	sky	$\checkmark$	[sky, morning, summer, sun, night]
she picked up the	ball	first	$\times$	[first, idea, news, story, best]
he put the food in his	mouth	hands	$\times$	[hands, kitchen, house, own, car]
the little girl was wearing a pretty	dress	black	$\sim$	[black, girl, dress, white, long]
the bird flew up into the	sky	air	$\sim$	[air, sky, ground, forest, water]

Table 28: AWD-LSTM: Baseline (

k=0

), all 100 Cloze prompts.

Prompt	Exp.	Pred.		Top-5
once upon a	time	a	$\times$	[a, home, family, in, -]
the sun was shining bright in the	sky	summer	$\times$	[summer, year, day, late, time]
she picked up the	ball	number	$\times$	[number, world, most, best, first]
he put the food in his	mouth	family	$\times$	[family, ,, book, death, own]
the little girl was wearing a pretty	dress	,	$\times$	[,, -, and, black, woman]
the bird flew up into the	sky	air	$\times$	[air, sea, ground, ocean, field]

Table 29: AWD-LSTM: Whole-Unit Removal ( weight_ih + weight_hh ,

k=0.2

, main-paper target), all 100 Cloze prompts.

Prompt	Exp.	Pred.		Top-5
once upon a	time	time	$\checkmark$	[time, ship, a, single, surface]
the sun was shining bright in the	sky	sky	$\checkmark$	[sky, sun, light, skies, dark]
she picked up the	ball	first	$\times$	[first, ”, last, next, most]
he put the food in his	mouth	house	$\times$	[house, car, mind, book, own]
the little girl was wearing a pretty	dress	little	$\times$	[little, black, girl, long, white]
the bird flew up into the	sky	air	$\sim$	[air, sky, desert, ground, sea]

Table 30: AWD-LSTM: Recurrent-Connection Pruning ( weight_hh ,

k=0.2

, supplementary control), all 100 Cloze prompts.

Prompt	Exp.	Pred.		Top-5
once upon a	time	life	$\times$	[life, live, film, day, history]
the sun was shining bright in the	sky	sky	$\checkmark$	[sky, black, day, life, days]
she picked up the	ball	idea	$\times$	[idea, film, girls, women, action]
he put the food in his	mouth	family	$\times$	[family, life, way, case, place]
the little girl was wearing a pretty	dress	girl	$\times$	[girl, black, -, woman, ,]
the bird flew up into the	sky	air	$\times$	[air, space, flight, city, black]

Table 31: AWD-LSTM: Input-Connection Pruning ( weight_ih ,

k=0.2

, supplementary control), all 100 Cloze prompts.

Prompt	Exp.	Pred.		Top-5
once upon a	time	year	$\sim$	[year, period, time, new, few]
the sun was shining bright in the	sky	middle	$\times$	[middle, same, face, first, second]
she picked up the	ball	first	$\times$	[first, ”, the, second, new]
he put the food in his	mouth	own	$\times$	[own, home, car, first, life]
the little girl was wearing a pretty	dress	long	$\times$	[long, little, large, high, a]
the bird flew up into the	sky	city	$\times$	[city, united, country, area, first]

Table 32: Small LSTM: Baseline (

k=0

), all 100 Cloze prompts.

Prompt	Exp.	Pred.		Top-5
once upon a	time	”	$\times$	[”, year, day, long, few]
the sun was shining bright in the	sky	first	$\times$	[first, world, same, united, middle]
she picked up the	ball	first	$\times$	[first, ”, the, second, same]
he put the food in his	mouth	own	$\times$	[own, first, life, home, time]
the little girl was wearing a pretty	dress	-	$\times$	[-, ,, a, ., the]
the bird flew up into the	sky	first	$\times$	[first, world, second, the, ”]

Table 33: Small LSTM: Whole-Unit Removal ( weight_ih + weight_hh ,

k=0.2

, main-paper target), all 100 Cloze prompts.

Prompt	Exp.	Pred.		Top-5
once upon a	time	new	$\times$	[new, few, ”, brief, single]
the sun was shining bright in the	sky	same	$\times$	[same, area, first, middle, house]
she picked up the	ball	end	$\times$	[end, the, ”, first, new]
he put the food in his	mouth	own	$\times$	[own, home, first, life, family]
the little girl was wearing a pretty	dress	long	$\times$	[long, large, -, much, a]
the bird flew up into the	sky	first	$\times$	[first, new, the, ”, end]

Table 34: Small LSTM: Recurrent-Connection Pruning ( weight_hh ,

k=0.2

, supplementary control), all 100 Cloze prompts.

Prompt	Exp.	Pred.		Top-5
once upon a	time	part	$\sim$	[part, time, ”, the, year]
the sun was shining bright in the	sky	the	$\times$	[the, and, year, world, season]
she picked up the	ball	the	$\times$	[the, a, to, ,, .]
he put the food in his	mouth	,	$\times$	[,, ., and, to, -]
the little girl was wearing a pretty	dress	-	$\times$	[-, ,, ., and, in]
the bird flew up into the	sky	the	$\times$	[the, world, -, ,, and]

Table 35: Small LSTM: Input-Connection Pruning ( weight_ih ,

k=0.2

, supplementary control), all 100 Cloze prompts.

Can Language Models Get "Aphasia" From Neuronal Ablation?

Abstract

Figures & tables

Causal Interventions in Language Models

Artificial Aphasias in Lesioned Language Models

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

Language Impairment as a Lens onto Large Language Models

Abstract

Causal Interventions in Language Models

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

Language Impairment as a Lens onto Large Language Models