Mechanistic Failure Modes of Weight Pruning in Transformer Language Models

Nabeelah Faumi; Deshan Sumanathilaka

cs.LGOct 27, 2026

Mechanistic Failure Modes of Weight Pruning in Transformer Language Models

Authors: Nabeelah Faumi, Deshan Sumanathilaka

Organizations: Informatics Institute of Technology, 57, Ramakrishna Road, Colombo 06, Sri Lanka

Abstract

Pruned language models are judged almost entirely by accuracy and perplexity, which assumes that preserved output behaviour implies preserved internal computation. We test this assumption with a factorial pruning study on GPT-2, crossing two importance criteria (magnitude, random), two component types (attention heads, MLP neurons) and three sparsity levels (3%, 50%, 70%). Using the fully characterised indirect object identification (IOI) circuit as ground truth, we measure circuit integrity with activation patching, path patching and direct logit attribution. We find two qualitatively distinct failure modes. Magnitude head pruning >= zeroes the circuit: output-facing heads are removed, and the model falls to near-chance. Magnitude neuron pruning at 30% inverts it: all 26 circuit heads remain intact, but Name Mover contributions flip sign, and the model systematically prefers the wrong name (17.7% accuracy). Perplexity ranks it as less damaged than random head pruning (74.6 vs 140.1). A replication of the Greater-Than circuit shows that the critical component type depends on the circuit's architecture. These results show that output metrics and structural checks can both miss mechanistic damage, motivating circuit-level causal analysis as a complementary evaluation of compressed language models.

Figures & tables

Method	Model	Mean LD	Acc (%)	Degrad. (%)
Base	GPT-2 Small (FP32)	$2.942$	$94.8$	,
Pruning 30%	Mag. Neurons	$-2.475$	$17.7$	$184.1$
	Rand. Neurons	$-0.896$	$35.6$	$130.5$
	Mag. Attn	$0.200$	$51.4$	$93.2$
	Rand. Attn	$-0.706$	$36.4$	$124.0$
Pruning 50%	Mag. Neurons	$-1.318$	$33.7$	$144.8$

Table 1: IOI task performance across all compressed models. Degradation

=(\mu_{\text{base}}-\mu_{\text{comp}})/\mu_{\text{base}}

; values

>100\%

indicate a negative mean logit difference.

Method	Model	S-INH	NM	NNM	BNM	Failure Mode
Pruning 30%	Mag. Attn	$17.6$	$9.2$	$\approx 0$	$-2.2$	Zeroing (partial)
	Rand. Attn	$50.2$	$-8.9$	$-11.5$	$-59.8$	Inversion (partial)
	Mag. Neu	$21.2$	$-30.8$	$-23.0$	$-119.9$	Inversion
	Rand. Neu	$26.8$	$-8.5$	$-3.7$	$-44.0$	Inversion (partial)
Pruning 50%	Mag. Attn	$-7.1$	$\approx 0$	$\approx 0$	$\approx 0$	Zeroing
	Rand. Attn	$10.0$	$-5.3$	$-5.4$	$-37.3$	Inversion (partial)

Table 2: DLA retention (%) by circuit class across compressed models. Negative values indicate sign inversion (the class actively harms correct IO prediction).

\approx 0

denotes exact zero from structurally eliminated heads.

Spar.	Model	PPL	Cos.	CKA	CKA (early)	CKA (late)
30%	Mag. Neu	$74.61$	$.856$	$.698$	$.946$	$.508$
	Rand. Neu	$102.09$	$.917$	$.825$	$.993$	$.553$
	Mag. Attn	$23.37$	$.980$	$.941$	$.998$	$.848$
	Rand. Attn	$140.10$	$.880$	$.801$	$.986$	$.648$
50%	Mag. Neu	$262.28$	$.813$	$.577$	$.900$	$.366$
	Rand. Neu	$300.49$	$.897$	$.791$	$.977$	$.520$

Table 3: TinyStories perplexity and layer-wise representational similarity to the base model (base perplexity 11.54). Early

=

L0–L3, Late

=

L8–L11.

Spar.	Model	Mean $\Delta GT$	Acc (%)	Degrad. (%)
30%	Mag. Neu	$-0.580\pm 0.643$	$0.2$	$120.3$
	Rand. Neu	$0.421\pm 1.286$	$33.8$	$85.3$
	Mag. Attn	$1.197\pm 1.046$	$66.1$	$58.1$
	Rand. Attn	$1.066\pm 1.701$	$49.9$	$62.6$
50%	Mag. Neu	$-0.369\pm 0.730$	$0.0$	$112.9$
	Rand. Neu	$-0.416\pm 0.349$	$0.0$	$114.6$

Table 4: GT logit difference, accuracy and degradation under pruning. Degradation

>100\%

indicates a negative mean logit difference. Base

\Delta GT=2.855

		YE $\to$ YD		YD $\to$ CMP		YD $\to$ OUT		CMP $\to$ OUT
Spar.	Model	Deg.	$d$	Deg.	$d$	Deg.	$d$	Deg.	$d$
30%	Mag. Neu	$49.9$	$-.55$	$-11.4$	$.06$	$22.6$	$-.17$	$54.4$	$-.67$
	Rand. Neu	$44.4$	$-.56$	$-18.8$	$.12$	$17.4$	$-.15$	$49.0$	$-.68$
	Mag. Attn	$38.4$	$-1.16$	$1.1$	$-.02$	$27.0$	$-.57$	$42.9$	$-1.49$
	Rand. Attn	$22.6$	$-.26$	$39.0$	$-.22$	$41.2$	$-.33$	$15.4$	$-.20$
50%	Mag. Neu	$62.2$	$-.71$	$19.4$	$-.11$	$43.9$	$-.34$	$65.3$	$-.82$

Table 5: GT residual path patching: mean degradation (%) and Cohen’s

d

. Negative degradation indicates increased causal potency in the compressed model.

Appendix figures & tables5 assets

Supplementary material from the paper’s appendix.

Appendix

Circuit Class	Class DLA	Top Head	Head DLA
Name Mover (NM)	$+3.576$	L9H9	$+2.247$
		L9H6	$+0.913$
		L10H0	$+0.557$
		L10H6	$+0.208$
Negative Name Mover (NNM)	$-2.094$	L10H7	(dominant NNM)
Backup Name Mover (BNM)	$+0.427$	L10H10	$+0.455$

Appendix

Table 6: Base GPT2 DLA scores by IOI circuit class and top individual heads. Positive scores contribute towards IO over S; negative towards S (correct for NNM).

		PT $\to$ IND		DT $\to$ S-INH		IND $\to$ NM		S-INH $\to$ NM
Spar.	Model	Deg.	$d$	Deg.	$d$	Deg.	$d$	Deg.	$d$
30%	Mag. Neu	$255.0\%$	$.108^{\checkmark}$	$158.0\%$	$.130^{\checkmark}$	$201.0\%$	$.086^{\checkmark}$	$\mathbf{117.1\%}$	$.284^{\checkmark}$
	Rand. Neu	$-96.4\%$	$-.053^{\checkmark}$	$-41.4\%$	$-.043$	$-182.4\%$	$-.100^{\checkmark}$	$35.4\%$	$.102^{\checkmark}$
	Mag. Attn	$66.5\%$	$.049^{\checkmark}$	$62.6\%$	$.088^{\checkmark}$	$36.6\%$	$.027$	$\mathbf{87.4\%}$	$.365^{\checkmark}$
	Rand. Attn	$201.3\%$	$.117^{\checkmark}$	$122.0\%$	$.137^{\checkmark}$	$143.2\%$	$.084^{\checkmark}$	$\mathbf{91.2\%}$	$.310^{\checkmark}$
50%	Mag. Neu	$139.8\%$	$.048^{\checkmark}$	$112.1\%$	$.075^{\checkmark}$	$111.2\%$	$.039$	$\mathbf{102.4\%}$	$.200^{\checkmark}$

Appendix

Table 7: Residual path patching: mean degradation (%) and paired Cohen’s

d

for four critical pathways (

n=2{,}000

per path per model). Bold entries indicate the largest effect per model. Entries marked

\checkmark

are significant (

p<0.05

, Bonferroni-corrected). Negative degradation indicates circuit reorganisation (canonical path carries proportionally more causal weight in isolation in the compressed model).

Spar.	Model	$\bar{H}$	Supp.	$f_{0}$	$\bar{s}$	Std $(s)$	$\Delta H$
	Base GPT-2	$2.013$	$11.38$	$.421$	$1.000$	–	$0.000$
30%	Mag. Neu	$2.186$	$13.47$	$.386$	$.775$	$.090$	$+0.173$
	Rand. Neu	$1.638$	$9.25$	$.521$	$.894$	$.095$	$-0.376$
	Mag. Attn	$2.707$	$28.43$	$.270$	$.795$	$.330$	$+0.694$
	Rand. Attn	$2.604$	$26.27$	$.284$	$.704$	$.252$	$+0.590$
50%	Mag. Neu	$2.340$	$15.92$	$.360$	$.706$	$.105$	$+0.326$

Appendix

Table 8: Attention pattern statistics under pruning, averaged over heads.

\bar{H}

: mean entropy (nats); Supp.: effective support

\exp(\bar{H})

;

f_{0}

: first-token (attention-sink) attention;

\bar{s}

: mean per-head cosine similarity to base (Std: standard deviation across heads);

\Delta H

: mean per-head entropy change (nats).

\bar{s}\approx 1

indicates routing identical to base.

Spar.	Model	PT	DT	IND	S-INH	NM	NNM	BNM	All $d$
30%	Mag. Neu	$.121$	$.140$	$.149$	$.314$	$.200$	$-.341$	$.140$	$.135$
	Rand. Neu	$-.063$	$-.037$	$-.045$	$.141$	$-.016$	$-.539$	$-.059$	$-.053$
	Mag. Attn	$.046$	$.099$	$.102$	$.331$	$.035$	$-.467$	$.041$	$.053$
	Rand. Attn	$.107$	$.143$	$.153$	$.324$	$.160$	$-.417$	$.123$	$.118$
50%	Mag. Neu	$.058$	$.084$	$.092$	$.230$	$.092$	$-.244$	$.057$	$.074$
	Mag. Attn	$.078$	$.134$	$.161$	$.414$	$.113$	$-.345$	$.083$	$.118$

Appendix

Table 9: Head-level activation patching: paired Cohen’s

d

by functional class. Positive

d

indicates degraded causal contribution; negative

d

indicates paradoxically larger patching effect in the compressed model (disinhibition). All overall effects significant (

p<0.001

, Bonferroni-corrected) unless marked

\dagger

(

p<0.05

, uncorrected).

Method	Model	L1 Error	L1 Valid	L6 Error	L6 Valid	Mean Surv. (L6)
Base	GPT-2 Small (FP32)	$0.44$	$\checkmark$	$0.80$	$\checkmark$	$0.018$
Mag. Attn	30%	$0.44$	$\checkmark$	$2.44$	$\times$	,
	50%	$0.44$	$\checkmark$	$1.28$	$\times$	,
	70%	$0.35$	$\checkmark$	$0.38$	$\checkmark$	$0.0136$
Mag. Neu	30%	$\mathbf{155.87}$	$\times$	$0.67$	$\checkmark$	,
	50%	$117.50$	$\times$	$5.27$	$\times$	,

Appendix

Table 10: SAE reconstruction errors (relative) and mean feature survival by layer across compressed models.

\checkmark

= valid (relative error

<1.0

);

\times

= invalid (activations outside SAE training distribution). Layer 11 is excluded: base reconstruction error

=2.98

exceeds the validity threshold. Mean survival is reported only for the highest-sparsity valid case per strategy.

LLM Pruning

Large language model pruning removes less useful weights or components to reduce memory and inference costs while preserving model capabilities.

See all 178 papers

Magnitude Profile Pruning: Calibration-Free Structured Attention Head Removal for Transformer Compression
Aug 11, 2026Kasun Dewage, Marianna Pensky, Heranga K. Rathnasekara +1LLM Compression Transformer Pruning
University of Central Florida Orlando, Florida, USA · Old Dominion University Norfolk, Virginia, USA
How Pruning Reshapes Features: Sparse Autoencoder Analysis of Weight-Pruned Language Models
Mar 26, 2026Hector Borobia, Elies Seguí-Mas, Guillermina Tormo-CarbóLLM Pruning Representation Geometry in Language Models
VRAIN – Valencian Research Institute for Artificial Intelligence, Universitat Politècnica de València, Valencia, Spain · Department of Economics and Social Sciences, Universitat Politècnica de València, Valencia, Spain · Department of Business Organisation, Universitat Politècnica de València, Valencia, Spain
Pruning via Causal Attribution Preserves Reasoning Performance in Large Language Models
Apr 27, 2026Amogh Sheth, Andrew Lin, Biruk Assefa +2LLM Pruning Self-Attention
Edison Academy Magnet School · Massachusetts Institute of Technology · State University of New York College at Plattsburgh +2

Aug 11, 2026J/K move · Enter open · S save

Magnitude Profile Pruning: Calibration-Free Structured Attention Head Removal for Transformer Compression

Kasun Dewage, Marianna Pensky, Heranga K. Rathnasekara, Suranadi De Silva

University of Central Florida Orlando, Florida, USA · Old Dominion University Norfolk, Virginia, USA

LLM Compression Transformer Pruning Attention Head Pruning Neural Network Pruning Transformer Attention

Open paper AskPDF arXiv

Structured pruning of attention heads provides a hardware-friendly way to compress Transformer language models. However, existing methods for measuring head-level importance require calibration data, gradient computation, or Hessian estimation. These requirements add extra overhead and make the methods depend on the data. Our work presents Magnitude Profile (MP) scoring, a training-free criterion for head importance that identifies dispensable heads through statistical outlier detection on weight row norms. Heads whose projection weights fall within the population bulk are pruned, while heads exhibiting outlier norms, which carry disproportionate representational capacity, are preserved. Our work further gives MP-G, a variant that handles Grouped Query Attention (GQA) by distributing shared key-value group scores across associated query heads. Across five models evaluated on WikiText-2 perplexity at 12.5%-50% head sparsity, MP-G achieves the best perplexity on OPT-6.7B at all sparsity levels (18.46 at 12.5%, 27.87 at 25%, 152.0 at 50%). MP-G also gives the best results on RoBERTa-large at 12.5% and 25% sparsity, with perplexity values of 7.27 and 10.28, outperforming calibration-dependent baselines including Wanda-Head, SparseGPT-Head, and Gradient-Head. It requires zero forward passes, calibration samples, or gradient computation. At 50% sparsity, head pruning yields up to 16% parameter reduction with 50% attention FLOP savings. Our results show that weight-only statistical scoring can match or outperform data-dependent methods for structured head pruning, providing a practical, zero-cost criterion for Transformer compression.

Abstract

LLM Pruning

Explore similar work

Magnitude Profile Pruning: Calibration-Free Structured Attention Head Removal for Transformer Compression

How Pruning Reshapes Features: Sparse Autoencoder Analysis of Weight-Pruned Language Models

Pruning via Causal Attribution Preserves Reasoning Performance in Large Language Models