pith. sign in

arxiv: 2606.24970 · v1 · pith:LXQ2DY2Knew · submitted 2026-06-23 · 💻 cs.LG

Don't Go Breaking My LLM: The Impact of Pruning Attention Layers on Explanation Faithfulness and Confidence Calibration

Pietro Tropeano , Maria Maistro , Tuukka Ruotsalo , Christina Lioma This is my paper

Pith reviewed 2026-06-26 01:04 UTC · model grok-4.3

classification 💻 cs.LG
keywords LLM pruningattention layersexplanation faithfulnessconfidence calibrationmodel interpretabilitymodel compressionlarge language modelsreliability
0
0 comments X

The pith

Pruning attention layers in LLMs often degrades explanation faithfulness and confidence calibration even when accuracy holds steady.

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

The paper tests the effects of removing attention layers from large language models to reduce their size and speed. It evaluates five models across eight datasets and finds that accuracy frequently survives the pruning, yet the faithfulness of generated explanations and the alignment between confidence scores and actual correctness commonly worsen. These two properties can shift markedly from one pruning level to the next even while accuracy stays flat, exposing a gap between what standard benchmarks measure and what users need for trustworthy outputs. A reader would care because many downstream uses rely on explanations that reflect the model's actual reasoning and on confidence values that can be trusted to flag uncertain predictions. The authors therefore argue that compression evaluations must track these additional dimensions.

Core claim

Removing up to one-third of the attention layers preserves most accuracy in the tested LLMs, but the faithfulness of explanations and the calibration of confidence scores frequently decline. These declines occur independently of accuracy changes and can vary substantially across pruning ratios, models, and datasets, revealing that accuracy and efficiency metrics alone do not capture the full impact on interpretability and reliability.

What carries the argument

The removal of attention layers from transformer-based LLMs, measured against faithfulness metrics for explanations and calibration metrics for confidence scores.

If this is right

  • Pruned models may generate explanations that do not match their internal decision process.
  • Confidence scores from pruned models may fail to indicate when predictions are likely to be wrong.
  • Accuracy figures alone cannot be relied upon to certify the quality of a compressed LLM.
  • Standard pruning evaluations should be expanded to include faithfulness and calibration checks.
  • The misalignment between accuracy and the other properties can differ by model and by dataset.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • Teams deploying pruned models in settings that require explanations may need additional post-pruning adjustments to restore faithfulness.
  • The same independence of accuracy from calibration could appear under other compression techniques such as quantization or distillation.
  • Safety evaluations for compressed LLMs should treat faithfulness and calibration as first-class requirements rather than optional add-ons.

Load-bearing premise

The observed effects depend on the assumption that the particular pruning strategy and the chosen faithfulness and calibration metrics are suitable and representative for the models and tasks examined.

What would settle it

An experiment showing that faithfulness and calibration scores remain stable or improve after the same attention-layer pruning on the same models and datasets would falsify the central claim.

Figures

Figures reproduced from arXiv: 2606.24970 by Christina Lioma, Maria Maistro, Pietro Tropeano, Tuukka Ruotsalo.

Figure 1
Figure 1. Figure 1: Importance scores of attention layers for all models. The importance scores are computed as 1 - [PITH_FULL_IMAGE:figures/full_fig_p004_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Accuracy of pruned and unpruned models (y axis) on eight different datasets, with varying numbers [PITH_FULL_IMAGE:figures/full_fig_p006_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Accuracy fluctuations (y-axis) [PITH_FULL_IMAGE:figures/full_fig_p007_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Overlap between the top 5, 10, and 20 LIME and Kernel SHAP features (y-axis) for pruned and [PITH_FULL_IMAGE:figures/full_fig_p008_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Relative change in LIME’s comprehensiveness and sufficiency, and in accuracy (y axis) between [PITH_FULL_IMAGE:figures/full_fig_p009_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: First row: ECE (y axis) for Mistral 7B and Llama-3 8B for different levels of attention layer [PITH_FULL_IMAGE:figures/full_fig_p011_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: Columns (left to right): average F1, Precision, Recall, and IoU between LIME attributions for Mistral 7B and human annotations (y axis), as a function of pruned attention layers (x axis). Dotted lines denote the unpruned model. For F1, Precision, and Recall, colours indicate the proportion of top features considered (K). Rows correspond to datasets: ARC Easy (top) and TweetEval (bottom). The fourth column … view at source ↗
Figure 8
Figure 8. Figure 8: Number of fluctuations (x-axis) in accuracy (first barplot) and ECE (second barplot) for combi [PITH_FULL_IMAGE:figures/full_fig_p024_8.png] view at source ↗
Figure 9
Figure 9. Figure 9: Number of fluctuations (x-axis) in comprehensiveness (first barplot) and sufficiency (second [PITH_FULL_IMAGE:figures/full_fig_p025_9.png] view at source ↗
Figure 10
Figure 10. Figure 10: Overlap between the top 5, 10, and 20 features identified by LIME (y-axis) for pruned and [PITH_FULL_IMAGE:figures/full_fig_p026_10.png] view at source ↗
Figure 11
Figure 11. Figure 11: Overlap between the top 5, 10, and 20 features identified by Kernel SHAP (y-axis) for pruned [PITH_FULL_IMAGE:figures/full_fig_p027_11.png] view at source ↗
Figure 12
Figure 12. Figure 12: Comprehensiveness (↑), sufficiency (↓), and accuracy (↑) of models (y axis) on QA tasks, using LIME, at varying amounts of removed attention layers (x axis). 28 [PITH_FULL_IMAGE:figures/full_fig_p028_12.png] view at source ↗
Figure 13
Figure 13. Figure 13: Comprehensiveness (↑), sufficiency (↓), and accuracy (↑) of models (y axis) on sentiment analysis tasks and RTE, using LIME, at varying amounts of removed attention layers (x axis). 29 [PITH_FULL_IMAGE:figures/full_fig_p029_13.png] view at source ↗
Figure 14
Figure 14. Figure 14: Comprehensiveness (↑), sufficiency (↓), and accuracy (↑) of models (y axis) on QA tasks, using Kernel SHAP, at varying amounts of removed attention layers (x axis). 30 [PITH_FULL_IMAGE:figures/full_fig_p030_14.png] view at source ↗
Figure 15
Figure 15. Figure 15: Comprehensiveness (↑), sufficiency (↓), and accuracy (↑) of models (y axis) on sentiment analysis tasks and RTE, using Kernel SHAP, at varying amounts of removed attention layers (x axis). 31 [PITH_FULL_IMAGE:figures/full_fig_p031_15.png] view at source ↗
Figure 16
Figure 16. Figure 16: Partial comprehensiveness (↑), sufficiency (↓), and accuracy (↑) of models (y axis) on QA tasks, using LIME, at varying amounts of removed attention layers (x axis). 32 [PITH_FULL_IMAGE:figures/full_fig_p032_16.png] view at source ↗
Figure 17
Figure 17. Figure 17: Partial comprehensiveness (↑), sufficiency (↓), and accuracy (↑) of models (y axis) on sentiment analysis and RTE tasks, using LIME, at varying amounts of removed attention layers (x axis). 33 [PITH_FULL_IMAGE:figures/full_fig_p033_17.png] view at source ↗
Figure 18
Figure 18. Figure 18: Partial comprehensiveness (↑), sufficiency (↓), and accuracy (↑) of models (y axis) on QA tasks, using Kernel SHAP, at varying amounts of removed attention layers (x axis). 34 [PITH_FULL_IMAGE:figures/full_fig_p034_18.png] view at source ↗
Figure 19
Figure 19. Figure 19: Partial comprehensiveness (↑), sufficiency (↓), and accuracy (↑) of models (y axis) on sentiment analysis and RTE tasks, using Kernel SHAP, at varying amounts of removed attention layers (x axis). 35 [PITH_FULL_IMAGE:figures/full_fig_p035_19.png] view at source ↗
Figure 20
Figure 20. Figure 20: Relative reduction in comprehensiveness and sufficiency using LIME, and accuracy (y axis) [PITH_FULL_IMAGE:figures/full_fig_p036_20.png] view at source ↗
Figure 21
Figure 21. Figure 21: Relative reduction in comprehensiveness and sufficiency using LIME, and accuracy (y axis) [PITH_FULL_IMAGE:figures/full_fig_p037_21.png] view at source ↗
Figure 22
Figure 22. Figure 22: Relative reduction in comprehensiveness and sufficiency using Kernel SHAP, and accuracy (y [PITH_FULL_IMAGE:figures/full_fig_p038_22.png] view at source ↗
Figure 23
Figure 23. Figure 23: Relative reduction in comprehensiveness and sufficiency using Kernel SHAP, and accuracy (y [PITH_FULL_IMAGE:figures/full_fig_p039_23.png] view at source ↗
Figure 24
Figure 24. Figure 24: F1 score (y-axis) between LIME attributions and human annotations as a function of pruned [PITH_FULL_IMAGE:figures/full_fig_p040_24.png] view at source ↗
Figure 25
Figure 25. Figure 25: F1 score (y-axis) between Kernel SHAP attributions and human annotations as a function of [PITH_FULL_IMAGE:figures/full_fig_p041_25.png] view at source ↗
Figure 26
Figure 26. Figure 26: Precision (y-axis) between LIME attributions and human annotations as a function of pruned [PITH_FULL_IMAGE:figures/full_fig_p042_26.png] view at source ↗
Figure 27
Figure 27. Figure 27: Precision (y-axis) between Kernel SHAP attributions and human annotations as a function of [PITH_FULL_IMAGE:figures/full_fig_p043_27.png] view at source ↗
Figure 28
Figure 28. Figure 28: Recall (y-axis) between LIME attributions and human annotations as a function of pruned [PITH_FULL_IMAGE:figures/full_fig_p044_28.png] view at source ↗
Figure 29
Figure 29. Figure 29: Recall (y-axis) between Kernel SHAP attributions and human annotations as a function of pruned [PITH_FULL_IMAGE:figures/full_fig_p045_29.png] view at source ↗
Figure 30
Figure 30. Figure 30: Intersection-over-Union (y-axis) between LIME attributions and human annotations as a function [PITH_FULL_IMAGE:figures/full_fig_p046_30.png] view at source ↗
Figure 31
Figure 31. Figure 31: Intersection-over-Union (y-axis) between Kernel SHAP attributions and human annotations as a [PITH_FULL_IMAGE:figures/full_fig_p047_31.png] view at source ↗
Figure 32
Figure 32. Figure 32: Area Under the Precision-Recall Curve (y-axis) between LIME attributions and human annota [PITH_FULL_IMAGE:figures/full_fig_p048_32.png] view at source ↗
Figure 33
Figure 33. Figure 33: Area Under the Precision-Recall Curve (y-axis) between Kernel SHAP attributions and human [PITH_FULL_IMAGE:figures/full_fig_p049_33.png] view at source ↗
Figure 34
Figure 34. Figure 34: Expected Calibration Error (y axis) for Mistral (first row), Llama-2 (second row), Llama-3.1 [PITH_FULL_IMAGE:figures/full_fig_p050_34.png] view at source ↗
Figure 35
Figure 35. Figure 35: Relative reduction in ECE and accuracy (y axis) for Mistral (first row), Llama-2 (second row), [PITH_FULL_IMAGE:figures/full_fig_p051_35.png] view at source ↗
read the original abstract

Pruning Large Language Models (LLMs) reduces memory and inference costs by removing parts of the network, producing smaller models that retain most of their accuracy. As attention layers are the most resource-intensive parts of LLMs, pruning them is a promising compression strategy. Prior work shows that up to 33% of attention layers can be pruned with minimal accuracy loss. Nevertheless, the impact of attention pruning on model interpretability, specifically faithfulness and confidence calibration, remains unstudied. To address this gap, we study how pruning attention layers affects explanation faithfulness and confidence calibration across five LLMs and eight datasets. While the pruned models often maintain high accuracy, we find that their faithfulness and calibration often degrade. Notably, faithfulness and calibration can fluctuate significantly, even when accuracy remains stable, highlighting a misalignment between model confidence, interpretability, and accuracy. Our findings suggest that layer pruning can affect LLMs' interpretability and reliability in ways not captured by accuracy and efficiency measures alone. We recommend including explainability and calibration metrics when evaluating pruned models.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

2 major / 2 minor

Summary. The paper conducts an empirical investigation into the effects of pruning attention layers (up to 33%) in five LLMs across eight datasets. It reports that pruned models frequently preserve high accuracy but exhibit degraded or fluctuating explanation faithfulness and confidence calibration, indicating a misalignment between accuracy, interpretability, and reliability. The authors conclude that evaluation of pruned LLMs should incorporate faithfulness and calibration metrics beyond accuracy and efficiency.

Significance. If the empirical findings are robust, the work is significant for highlighting that model compression via attention pruning can decouple accuracy from explanation quality and calibration in ways not captured by standard metrics. This could influence evaluation practices in LLM compression research by emphasizing the need for multi-faceted assessments of pruned models.

major comments (2)
  1. [§3 and §4] §3 (Experimental Setup) and §4 (Results): The faithfulness metrics are applied directly to pruned models without reported validation for invariance to attention-layer removal. Since many standard faithfulness metrics (sufficiency, comprehensiveness, or attention-weight based) depend on the very components removed by pruning, the observed degradations may be artifacts of metric sensitivity rather than substantive changes in explanation quality; this directly undermines the central claim of misalignment.
  2. [§4.2] §4.2 (Faithfulness and Calibration Results): The reported fluctuations in faithfulness/calibration while accuracy remains stable lack accompanying statistical controls (e.g., significance testing across the 5×8 model-dataset combinations or ablation on pruning ratios), making it unclear whether the fluctuations exceed noise or multiple-comparison artifacts.
minor comments (2)
  1. [Abstract] The abstract and introduction would benefit from explicit definitions or citations for the specific faithfulness and calibration metrics employed, even if expanded in the methods.
  2. [Tables in §4] Tables reporting per-model/per-dataset results should include error bars or variance measures to support claims of 'significant fluctuation'.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for their constructive feedback, which helps improve the robustness of our empirical analysis. We address each major comment below.

read point-by-point responses

  1. Referee: [§3 and §4] §3 (Experimental Setup) and §4 (Results): The faithfulness metrics are applied directly to pruned models without reported validation for invariance to attention-layer removal. Since many standard faithfulness metrics (sufficiency, comprehensiveness, or attention-weight based) depend on the very components removed by pruning, the observed degradations may be artifacts of metric sensitivity rather than substantive changes in explanation quality; this directly undermines the central claim of misalignment.

    Authors: We appreciate this concern about potential metric sensitivity. Our primary faithfulness metrics are perturbation-based (sufficiency and comprehensiveness), which evaluate explanation quality via input feature removal and are independent of internal attention layers; thus they remain valid post-pruning. Any attention-weight metrics are secondary. To further address the point, we will add a validation subsection in the revised §3 showing that these metrics yield stable rankings on a subset of pruned vs. original models, confirming the degradations are not artifacts. revision: yes

  2. Referee: [§4.2] §4.2 (Faithfulness and Calibration Results): The reported fluctuations in faithfulness/calibration while accuracy remains stable lack accompanying statistical controls (e.g., significance testing across the 5×8 model-dataset combinations or ablation on pruning ratios), making it unclear whether the fluctuations exceed noise or multiple-comparison artifacts.

    Authors: We agree that additional statistical controls would strengthen the results. In the revision we will add paired significance tests (e.g., Wilcoxon signed-rank) across all 5×8 combinations, report p-values with multiple-comparison correction, and include an ablation table for pruning ratios (10%, 20%, 33%) with error bars to demonstrate that observed fluctuations exceed noise levels. revision: yes

Circularity Check

0 steps flagged

Empirical study reports observations with no derivation chain

full rationale

The paper is an empirical investigation that measures the effects of attention-layer pruning on accuracy, faithfulness, and calibration across LLMs and datasets. It states prior results via citation and reports direct experimental outcomes without equations, fitted parameters renamed as predictions, self-definitional constructs, or any load-bearing self-citation chains. No derivation is claimed or present, so the findings rest on external measurements rather than internal reduction to inputs.

Axiom & Free-Parameter Ledger

0 free parameters · 1 axioms · 0 invented entities

The work is purely empirical with no new parameters, axioms beyond standard ones, or invented entities mentioned in the abstract.

axioms (1)
  • domain assumption Standard machine learning evaluation practices apply to the faithfulness and calibration metrics used.
    The abstract relies on these metrics being meaningful without defining them.

pith-pipeline@v0.9.1-grok · 5724 in / 1104 out tokens · 40144 ms · 2026-06-26T01:04:29.111336+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

61 extracted references · 24 canonical work pages

  1. [1]

    Slicegpt: Compress large language models by deleting rows and columns

    Saleh Ashkboos, Maximilian Croci, Marcelo Gennari do Nascimento, Torsten Hoefler, and James Hensman. Slicegpt: Compress large language models by deleting rows and columns. In B. Kim, Y. Yue, S. Chaudhuri, K. Fragkiadaki, M. Khan, and Y. Sun (eds.), International Conference on Learning Representations, volume 2024, pp.\ 11682--11701, 2024. URL https://proc...

    2024

  2. [2]

    Bach, Victor Sanh, Zheng-Xin Yong, Albert Webson, Colin Raffel, Nihal V

    Stephen H. Bach, Victor Sanh, Zheng-Xin Yong, Albert Webson, Colin Raffel, Nihal V. Nayak, Abheesht Sharma, Taewoon Kim, M Saiful Bari, Thibault Fevry, Zaid Alyafeai, Manan Dey, Andrea Santilli, Zhiqing Sun, Srulik Ben-David, Canwen Xu, Gunjan Chhablani, Han Wang, Jason Alan Fries, Maged S. Al-shaibani, Shanya Sharma, Urmish Thakker, Khalid Almubarak, Xia...

    work page doi:10.18653/v1/2022.acl-demo.9 2022

  3. [3]

    T weet E val: Unified Benchmark and Comparative Evaluation for Tweet Classification

    Francesco Barbieri, Jose Camacho-Collados, Luis Espinosa Anke, and Leonardo Neves. T weet E val: Unified benchmark and comparative evaluation for tweet classification. In Trevor Cohn, Yulan He, and Yang Liu (eds.), Findings of the Association for Computational Linguistics: EMNLP 2020, pp.\ 1644--1650, Online, November 2020. Association for Computational L...

    work page doi:10.18653/v1/2020.findings-emnlp.148 2020

  4. [4]

    The disagreement problem in faithfulness metrics

    Brian Barr, Noah Fatsi, Leif Hancox-Li, Peter Richter, and Daniel Proano. The disagreement problem in faithfulness metrics. In XAI in Action: Past, Present, and Future Applications, 2023. URL https://openreview.net/forum?id=KPtW2SU0my

    2023

  5. [5]

    The fifth pascal recognizing textual entailment challenge

    Luisa Bentivogli, Peter Clark, Ido Dagan, and Danilo Giampiccolo. The fifth pascal recognizing textual entailment challenge. TAC, 7 0 (8): 0 1, 2009. URL https://hdl.handle.net/11582/5351

    2009

  6. [6]

    A comparative study of faithfulness metrics for model interpretability methods

    Chun Sik Chan, Huanqi Kong, and Liang Guanqing. A comparative study of faithfulness metrics for model interpretability methods. In Smaranda Muresan, Preslav Nakov, and Aline Villavicencio (eds.), Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pp.\ 5029--5038, Dublin, Ireland, May 2022. Asso...

    work page doi:10.18653/v1/2022.acl-long.345 2022

  7. [7]

    Investigating hallucinations in pruned large language models for abstractive summarization

    George Chrysostomou, Zhixue Zhao, Miles Williams, and Nikolaos Aletras. Investigating hallucinations in pruned large language models for abstractive summarization. Transactions of the Association for Computational Linguistics, 12: 0 1163--1181, 2024. doi:10.1162/tacl_a_00695. URL https://aclanthology.org/2024.tacl-1.64/

    work page doi:10.1162/tacl_a_00695 2024

  8. [8]

    B ool Q : Exploring the Surprising Difficulty of Natural Yes/No Questions

    Christopher Clark, Kenton Lee, Ming-Wei Chang, Tom Kwiatkowski, Michael Collins, and Kristina Toutanova. B ool Q : Exploring the surprising difficulty of natural yes/no questions. In Jill Burstein, Christy Doran, and Thamar Solorio (eds.), Proceedings of the 2019 Conference of the North A merican Chapter of the Association for Computational Linguistics: H...

    work page doi:10.18653/v1/n19-1300 2019

  9. [9]

    Think you have solved question answering? try arc, the ai2 reasoning challenge, 2018

    Peter Clark, Isaac Cowhey, Oren Etzioni, Tushar Khot, Ashish Sabharwal, Carissa Schoenick, and Oyvind Tafjord. Think you have solved question answering? try arc, the ai2 reasoning challenge, 2018. URL https://arxiv.org/abs/1803.05457

    Pith/arXiv arXiv 2018

  10. [10]

    The pascal recognising textual entailment challenge

    Ido Dagan, Oren Glickman, and Bernardo Magnini. The pascal recognising textual entailment challenge. In Joaquin Qui \ n onero-Candela, Ido Dagan, Bernardo Magnini, and Florence d'Alch \'e Buc (eds.), Machine Learning Challenges. Evaluating Predictive Uncertainty, Visual Object Classification, and Recognising Tectual Entailment, pp.\ 177--190, Berlin, Heid...

    work page doi:10.1007/11736790_9 2006

  11. [11]

    Jay DeYoung, Sarthak Jain, Nazneen Fatema Rajani, Eric Lehman, Caiming Xiong, Richard Socher, and Byron C. Wallace. ERASER : A benchmark to evaluate rationalized NLP models. In Dan Jurafsky, Joyce Chai, Natalie Schluter, and Joel Tetreault (eds.), Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pp.\ 4443--4458, Onl...

    work page doi:10.18653/v1/2020.acl-main.408 2020

  12. [12]

    An unsupervised approach to achieve supervised-level explainability in healthcare records

    Joakim Edin, Maria Maistro, Lars Maal e, Lasse Borgholt, Jakob Drachmann Havtorn, and Tuukka Ruotsalo. An unsupervised approach to achieve supervised-level explainability in healthcare records. In Yaser Al-Onaizan, Mohit Bansal, and Yun-Nung Chen (eds.), Proceedings of the 2024 Conference on Empirical Methods in Natural Language Processing, pp.\ 4869--489...

    work page doi:10.18653/v1/2024.emnlp-main.280 2024

  13. [13]

    S parse GPT : Massive language models can be accurately pruned in one-shot

    Elias Frantar and Dan Alistarh. S parse GPT : Massive language models can be accurately pruned in one-shot. In Andreas Krause, Emma Brunskill, Kyunghyun Cho, Barbara Engelhardt, Sivan Sabato, and Jonathan Scarlett (eds.), Proceedings of the 40th International Conference on Machine Learning, volume 202 of Proceedings of Machine Learning Research, pp.\ 1032...

    2023

  14. [14]

    Pruning weights but not truth: Safeguarding truthfulness while pruning LLM s

    Yao Fu, Runchao Li, Xianxuan Long, Haotian Yu, Xiaotian Han, Yu Yin, and Pan Li. Pruning weights but not truth: Safeguarding truthfulness while pruning LLM s. In Christos Christodoulopoulos, Tanmoy Chakraborty, Carolyn Rose, and Violet Peng (eds.), Findings of the Association for Computational Linguistics: EMNLP 2025, pp.\ 20750--20768, Suzhou, China, Nov...

    work page doi:10.18653/v1/2025.findings-emnlp.1130 2025

  15. [15]

    The language model evaluation harness, 07 2024

    Leo Gao, Jonathan Tow, Baber Abbasi, Stella Biderman, Sid Black, Anthony DiPofi, Charles Foster, Laurence Golding, Jeffrey Hsu, Alain Le Noac'h, Haonan Li, Kyle McDonell, Niklas Muennighoff, Chris Ociepa, Jason Phang, Laria Reynolds, Hailey Schoelkopf, Aviya Skowron, Lintang Sutawika, Eric Tang, Anish Thite, Ben Wang, Kevin Wang, and Andy Zou. The languag...

    arXiv 2024

  16. [16]

    The third PASCAL recognizing textual entailment challenge

    Danilo Giampiccolo, Bernardo Magnini, Ido Dagan, and Bill Dolan. The third PASCAL recognizing textual entailment challenge. In Satoshi Sekine, Kentaro Inui, Ido Dagan, Bill Dolan, Danilo Giampiccolo, and Bernardo Magnini (eds.), Proceedings of the ACL - PASCAL Workshop on Textual Entailment and Paraphrasing , pp.\ 1--9, Prague, June 2007. Association for ...

    2007

  17. [17]

    Compressed but compromised? a study of jailbreaking in compressed LLM s

    Satya Sai Srinath Namburi GNVV, Alex James Boyd, and Andrew Warrington. Compressed but compromised? a study of jailbreaking in compressed LLM s. In Lock-LLM Workshop: Prevent Unauthorized Knowledge Use from Large Language Models, 2025. URL https://openreview.net/forum?id=OkNfb8SmLh

    2025

  18. [18]

    Andrey Gromov, Kushal Tirumala, Hassan Shapourian, Paolo Glorioso, and Daniel A. Roberts. The unreasonable ineffectiveness of the deeper layers. In Y. Yue, A. Garg, N. Peng, F. Sha, and R. Yu (eds.), International Conference on Learning Representations, volume 2025, pp.\ 81906--81920, 2025. URL https://proceedings.iclr.cc/paper_files/paper/2025/file/cbabc...

    2025

  19. [19]

    Weinberger

    Chuan Guo, Geoff Pleiss, Yu Sun, and Kilian Q. Weinberger. On calibration of modern neural networks. In Doina Precup and Yee Whye Teh (eds.), Proceedings of the 34th International Conference on Machine Learning, volume 70 of Proceedings of Machine Learning Research, pp.\ 1321--1330. PMLR, 06--11 Aug 2017. URL https://proceedings.mlr.press/v70/guo17a.html

    2017

  20. [20]

    The second pascal recognising textual entailment challenge

    R Bar Haim, Ido Dagan, Bill Dolan, Lisa Ferro, Danilo Giampiccolo, Bernardo Magnini, and Idan Szpektor. The second pascal recognising textual entailment challenge. In Proceedings of the Second PASCAL Challenges Workshop on Recognising Textual Entailment, volume 7, pp.\ 785--794, 2006

    2006

  21. [21]

    Pruning for protection: Increasing jailbreak resistance in aligned LLM s without fine-tuning

    Adib Hasan, Ileana Rugina, and Alex Wang. Pruning for protection: Increasing jailbreak resistance in aligned LLM s without fine-tuning. In Yonatan Belinkov, Najoung Kim, Jaap Jumelet, Hosein Mohebbi, Aaron Mueller, and Hanjie Chen (eds.), Proceedings of the 7th BlackboxNLP Workshop: Analyzing and Interpreting Neural Networks for NLP, pp.\ 417--430, Miami,...

    work page doi:10.18653/v1/2024.blackboxnlp-1.26 2024

  22. [22]

    Does BERT learn as humans perceive? understanding linguistic styles through lexica

    Shirley Anugrah Hayati, Dongyeop Kang, and Lyle Ungar. Does BERT learn as humans perceive? understanding linguistic styles through lexica. In Marie-Francine Moens, Xuanjing Huang, Lucia Specia, and Scott Wen-tau Yih (eds.), Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing, pp.\ 6323--6331, Online and Punta Cana, Domin...

    work page doi:10.18653/v1/2021.emnlp-main.510 2021

  23. [23]

    Uncovering the redundancy in transformers via a unified study of layer dropping

    Shwai He, Guoheng Sun, Zheyu Shen, and Ang Li. Uncovering the redundancy in transformers via a unified study of layer dropping. Transactions on Machine Learning Research, 2026. ISSN 2835-8856. URL https://openreview.net/forum?id=1I7PCbOPfe

    2026

  24. [24]

    Aligning \ ai \ with shared human values

    Dan Hendrycks, Collin Burns, Steven Basart, Andrew Critch, Jerry Li, Dawn Song, and Jacob Steinhardt. Aligning \ ai \ with shared human values. In International Conference on Learning Representations, 2021 a . URL https://openreview.net/forum?id=dNy_RKzJacY

    2021

  25. [25]

    Measuring massive multitask language understanding

    Dan Hendrycks, Collin Burns, Steven Basart, Andy Zou, Mantas Mazeika, Dawn Song, and Jacob Steinhardt. Measuring massive multitask language understanding. In International Conference on Learning Representations, 2021 b . URL https://openreview.net/forum?id=d7KBjmI3GmQ

    2021

  26. [26]

    Characterising bias in compressed models, 2020

    Sara Hooker, Nyalleng Moorosi, Gregory Clark, Samy Bengio, and Emily Denton. Characterising bias in compressed models, 2020. URL https://arxiv.org/abs/2010.03058

    arXiv 2020

  27. [27]

    Fasp: Fast and accurate structured pruning of large language models, 2025

    Hanyu Hu, Pengxiang Zhao, Ping Li, Yi Zheng, Zhefeng Wang, and Xiaoming Yuan. Fasp: Fast and accurate structured pruning of large language models, 2025. URL https://arxiv.org/abs/2501.09412

    arXiv 2025

  28. [28]

    Alon Jacovi and Yoav Goldberg. Towards faithfully interpretable NLP systems: How should we define and evaluate faithfulness? In Dan Jurafsky, Joyce Chai, Natalie Schluter, and Joel Tetreault (eds.), Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pp.\ 4198--4205, Online, July 2020. Association for Computational Lin...

    work page doi:10.18653/v1/2020.acl-main.386 2020

  29. [29]

    Albert Q. Jiang, Alexandre Sablayrolles, Arthur Mensch, Chris Bamford, Devendra Singh Chaplot, Diego de las Casas, Florian Bressand, Gianna Lengyel, Guillaume Lample, Lucile Saulnier, Lélio Renard Lavaud, Marie-Anne Lachaux, Pierre Stock, Teven Le Scao, Thibaut Lavril, Thomas Wang, Timothée Lacroix, and William El Sayed. Mistral 7b, 2023. URL https://arxi...

    Pith/arXiv arXiv 2023

  30. [30]

    The cost of down-scaling language models: Fact recall deteriorates before in-context learning, 2023

    Tian Jin, Nolan Clement, Xin Dong, Vaishnavh Nagarajan, Michael Carbin, Jonathan Ragan-Kelley, and Gintare Karolina Dziugaite. The cost of down-scaling language models: Fact recall deteriorates before in-context learning, 2023. URL https://arxiv.org/abs/2310.04680

    Pith/arXiv arXiv 2023

  31. [31]

    Logic traps in evaluating attribution scores

    Yiming Ju, Yuanzhe Zhang, Zhao Yang, Zhongtao Jiang, Kang Liu, and Jun Zhao. Logic traps in evaluating attribution scores. In Smaranda Muresan, Preslav Nakov, and Aline Villavicencio (eds.), Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pp.\ 5911--5922, Dublin, Ireland, May 2022. Associati...

    work page doi:10.18653/v1/2022.acl-long.407 2022

  32. [32]

    Shortened llama: Depth pruning for large language models with comparison of retraining methods, 2024

    Bo-Kyeong Kim, Geonmin Kim, Tae-Ho Kim, Thibault Castells, Shinkook Choi, Junho Shin, and Hyoung-Kyu Song. Shortened llama: Depth pruning for large language models with comparison of retraining methods, 2024. URL https://arxiv.org/abs/2402.02834

    arXiv 2024

  33. [33]

    The impact of inference acceleration on bias of LLM s

    Elisabeth Kirsten, Ivan Habernal, Vedant Nanda, and Muhammad Bilal Zafar. The impact of inference acceleration on bias of LLM s. In Luis Chiruzzo, Alan Ritter, and Lu Wang (eds.), Proceedings of the 2025 Conference of the Nations of the Americas Chapter of the Association for Computational Linguistics: Human Language Technologies (Volume 1: Long Papers), ...

    2025

  34. [34]

    Pruning filters for efficient convnets

    Hao Li, Asim Kadav, Igor Durdanovic, Hanan Samet, and Hans Peter Graf. Pruning filters for efficient convnets. In International Conference on Learning Representations, 2017. URL https://openreview.net/forum?id=rJqFGTslg

    2017

  35. [35]

    A unified approach to interpreting model predictions

    Scott M Lundberg and Su-In Lee. A unified approach to interpreting model predictions. Advances in neural information processing systems, 30, 2017

    2017

  36. [36]

    Towards faithful model explanation in NLP : A survey

    Qing Lyu, Marianna Apidianaki, and Chris Callison-Burch. Towards faithful model explanation in NLP : A survey. Computational Linguistics, 50 0 (2): 0 657--723, June 2024. doi:10.1162/coli_a_00511. URL https://aclanthology.org/2024.cl-2.6/

    work page doi:10.1162/coli_a_00511 2024

  37. [37]

    Llm-pruner: On the structural pruning of large language models

    Xinyin Ma, Gongfan Fang, and Xinchao Wang. Llm-pruner: On the structural pruning of large language models. In A. Oh, T. Naumann, A. Globerson, K. Saenko, M. Hardt, and S. Levine (eds.), Advances in Neural Information Processing Systems, volume 36, pp.\ 21702--21720. Curran Associates, Inc., 2023. URL https://proceedings.neurips.cc/paper_files/paper/2023/f...

    arXiv 2023

  38. [38]

    Shortgpt: Layers in large language models are more redundant than you expect, 2024

    Xin Men, Mingyu Xu, Qingyu Zhang, Bingning Wang, Hongyu Lin, Yaojie Lu, Xianpei Han, and Weipeng Chen. Shortgpt: Layers in large language models are more redundant than you expect, 2024. URL https://arxiv.org/abs/2403.03853

    arXiv 2024

  39. [39]

    Are sixteen heads really better than one? In H

    Paul Michel, Omer Levy, and Graham Neubig. Are sixteen heads really better than one? In H. Wallach, H. Larochelle, A. Beygelzimer, F. d Alch\' e -Buc, E. Fox, and R. Garnett (eds.), Advances in Neural Information Processing Systems, volume 32. Curran Associates, Inc., 2019. URL https://proceedings.neurips.cc/paper_files/paper/2019/file/2c601ad9d2ff9bc8b28...

    2019

  40. [40]

    Can a Suit of Armor Conduct Electricity? A New Dataset for Open Book Question Answering

    Todor Mihaylov, Peter Clark, Tushar Khot, and Ashish Sabharwal. Can a suit of armor conduct electricity? a new dataset for open book question answering. In Ellen Riloff, David Chiang, Julia Hockenmaier, and Jun ' ichi Tsujii (eds.), Proceedings of the 2018 Conference on Empirical Methods in Natural Language Processing, pp.\ 2381--2391, Brussels, Belgium, ...

    work page doi:10.18653/v1/d18-1260 2018

  41. [41]

    Cooper, and Milos Hauskrecht

    Mahdi Pakdaman Naeini, Gregory F. Cooper, and Milos Hauskrecht. Obtaining well calibrated probabilities using bayesian binning. In Proceedings of the Twenty-Ninth AAAI Conference on Artificial Intelligence, AAAI'15, pp.\ 2901–2907. AAAI Press, 2015. ISBN 0262511290. URL https://doi.org/10.1609/aaai.v29i1.9602

    work page doi:10.1609/aaai.v29i1.9602 2015

  42. [42]

    Seeing Stars: Exploiting Class Relationships for Sentiment Categorization with Respect to Rating Scales

    Bo Pang and Lillian Lee. Seeing stars: Exploiting class relationships for sentiment categorization with respect to rating scales. In Kevin Knight, Hwee Tou Ng, and Kemal Oflazer (eds.), Proceedings of the 43rd Annual Meeting of the Association for Computational Linguistics ( ACL ' 05) , pp.\ 115--124, Ann Arbor, Michigan, June 2005. Association for Comput...

    work page doi:10.3115/1219840.1219855 2005

  43. [43]

    Colin Raffel, Noam Shazeer, Adam Roberts, Katherine Lee, Sharan Narang, Michael Matena, Yanqi Zhou, Wei Li, and Peter J. Liu. Exploring the limits of transfer learning with a unified text-to-text transformer. Journal of Machine Learning Research, 21 0 (140): 0 1--67, 2020. URL http://jmlr.org/papers/v21/20-074.html

    2020

  44. [44]

    A comparative study on the impact of model compression techniques on fairness in language models

    Krithika Ramesh, Arnav Chavan, Shrey Pandit, and Sunayana Sitaram. A comparative study on the impact of model compression techniques on fairness in language models. In Anna Rogers, Jordan Boyd-Graber, and Naoaki Okazaki (eds.), Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pp.\ 15762--1578...

    work page doi:10.18653/v1/2023.acl-long.878 2023

  45. [45]

    and Guestrin C

    Marco Ribeiro, Sameer Singh, and Carlos Guestrin. why should I trust you? : Explaining the predictions of any classifier. In John DeNero, Mark Finlayson, and Sravana Reddy (eds.), Proceedings of the 2016 Conference of the North A merican Chapter of the Association for Computational Linguistics: Demonstrations , pp.\ 97--101, San Diego, California, June 20...

    work page doi:10.18653/v1/n16-3020 2016

  46. [46]

    S em E val-2017 task 4: Sentiment analysis in T witter

    Sara Rosenthal, Noura Farra, and Preslav Nakov. S em E val-2017 task 4: Sentiment analysis in T witter. In Steven Bethard, Marine Carpuat, Marianna Apidianaki, Saif M. Mohammad, Daniel Cer, and David Jurgens (eds.), Proceedings of the 11th International Workshop on Semantic Evaluation ( S em E val-2017) , pp.\ 502--518, Vancouver, Canada, August 2017. Ass...

    work page doi:10.18653/v1/s17-2088 2017

  47. [47]

    Sofia Serrano and Noah A. Smith. Is attention interpretable? In Anna Korhonen, David Traum, and Llu \'i s M \`a rquez (eds.), Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics, pp.\ 2931--2951, Florence, Italy, July 2019. Association for Computational Linguistics. doi:10.18653/v1/P19-1282. URL https://aclanthology.org...

    work page doi:10.18653/v1/p19-1282 2019

  48. [48]

    Who reasons in the large language models? In D

    Jie Shao and Jianxin Wu. Who reasons in the large language models? In D. Belgrave, C. Zhang, H. Lin, R. Pascanu, P. Koniusz, M. Ghassemi, and N. Chen (eds.), Advances in Neural Information Processing Systems, volume 38, pp.\ 113087--113108. Curran Associates, Inc., 2025. URL https://proceedings.neurips.cc/paper_files/paper/2025/file/a40462acc6959034c6aa6d...

    2025

  49. [49]

    A deeper look at depth pruning of LLM s

    Shoaib Ahmed Siddiqui, Xin Dong, Greg Heinrich, Thomas Breuel, Jan Kautz, David Krueger, and Pavlo Molchanov. A deeper look at depth pruning of LLM s. In ICML 2024 Workshop on Theoretical Foundations of Foundation Models, 2024. URL https://openreview.net/forum?id=9B7ayWclwN

    2024

  50. [50]

    Manning, Andrew Ng, and Christopher Potts

    Richard Socher, Alex Perelygin, Jean Wu, Jason Chuang, Christopher D. Manning, Andrew Ng, and Christopher Potts. Recursive deep models for semantic compositionality over a sentiment treebank. In David Yarowsky, Timothy Baldwin, Anna Korhonen, Karen Livescu, and Steven Bethard (eds.), Proceedings of the 2013 Conference on Empirical Methods in Natural Langu...

    2013

  51. [51]

    SLEB : Streamlining LLM s through redundancy verification and elimination of transformer blocks

    Jiwon Song, Kyungseok Oh, Taesu Kim, Hyungjun Kim, Yulhwa Kim, and jae-joon kim. SLEB : Streamlining LLM s through redundancy verification and elimination of transformer blocks. In Forty-first International Conference on Machine Learning, 2024. URL https://openreview.net/forum?id=fuX4hyLPmO

    2024

  52. [52]

    A simple and effective pruning approach for large language models

    Mingjie Sun, Zhuang Liu, Anna Bair, and J Zico Kolter. A simple and effective pruning approach for large language models. In The Twelfth International Conference on Learning Representations, 2024. URL https://openreview.net/forum?id=PxoFut3dWW

    2024

  53. [53]

    The llama 3 herd of models, 2024

    The Llama Team. The llama 3 herd of models, 2024. URL https://arxiv.org/abs/2407.21783

    Pith/arXiv arXiv 2024

  54. [54]

    Llama 2: Open foundation and fine-tuned chat models

    Hugo Touvron, Louis Martin, Kevin Stone, Peter Albert, Amjad Almahairi, Yasmine Babaei, Nikolay Bashlykov, Soumya Batra, Prajjwal Bhargava, Shruti Bhosale, et al. Llama 2: Open foundation and fine-tuned chat models. arXiv preprint arXiv:2307.09288, 2023. URL https://arxiv.org/abs/2307.09288

    Pith/arXiv arXiv 2023

  55. [55]

    GLUE : A Multi-Task Benchmark and Analysis Platform for Natural Language Understanding

    Alex Wang, Amanpreet Singh, Julian Michael, Felix Hill, Omer Levy, and Samuel Bowman. GLUE : A multi-task benchmark and analysis platform for natural language understanding. In Tal Linzen, Grzegorz Chrupa a, and Afra Alishahi (eds.), Proceedings of the 2018 EMNLP Workshop B lackbox NLP : Analyzing and Interpreting Neural Networks for NLP , pp.\ 353--355, ...

    work page doi:10.18653/v1/w18-5446 2018

  56. [56]

    Sheared LL a MA : Accelerating language model pre-training via structured pruning

    Mengzhou Xia, Tianyu Gao, Zhiyuan Zeng, and Danqi Chen. Sheared LL a MA : Accelerating language model pre-training via structured pruning. In The Twelfth International Conference on Learning Representations, 2024. URL https://openreview.net/forum?id=09iOdaeOzp

    2024

  57. [57]

    Beyond perplexity: Multi-dimensional safety evaluation of LLM compression

    Zhichao Xu, Ashim Gupta, Tao Li, Oliver Bentham, and Vivek Srikumar. Beyond perplexity: Multi-dimensional safety evaluation of LLM compression. In Yaser Al-Onaizan, Mohit Bansal, and Yun-Nung Chen (eds.), Findings of the Association for Computational Linguistics: EMNLP 2024, pp.\ 15359--15396, Miami, Florida, USA, November 2024. Association for Computatio...

    work page doi:10.18653/v1/2024.findings-emnlp.901 2024

  58. [58]

    Qwen3 technical report

    An Yang, Anfeng Li, Baosong Yang, Beichen Zhang, Binyuan Hui, Bo Zheng, Bowen Yu, Chang Gao, Chengen Huang, Chenxu Lv, et al. Qwen3 technical report. arXiv preprint arXiv:2505.09388, 2025. URL https://arxiv.org/abs/2505.09388

    Pith/arXiv arXiv 2025

  59. [60]

    Investigating layer importance in large language models

    Yang Zhang, Yanfei Dong, and Kenji Kawaguchi. Investigating layer importance in large language models. In Yonatan Belinkov, Najoung Kim, Jaap Jumelet, Hosein Mohebbi, Aaron Mueller, and Hanjie Chen (eds.), Proceedings of the 7th BlackboxNLP Workshop: Analyzing and Interpreting Neural Networks for NLP, pp.\ 469--479, Miami, Florida, US, November 2024 b . A...

    work page doi:10.18653/v1/2024.blackboxnlp-1.29 2024

  60. [61]

    Finercut: Finer-grained interpretable layer pruning for large language models, 2024 c

    Yang Zhang, Yawei Li, Xinpeng Wang, Qianli Shen, Barbara Plank, Bernd Bischl, Mina Rezaei, and Kenji Kawaguchi. Finercut: Finer-grained interpretable layer pruning for large language models, 2024 c . URL https://arxiv.org/abs/2405.18218

    arXiv 2024

  61. [62]

    Plug-and-play: An efficient post-training pruning method for large language models

    Yingtao Zhang, Haoli Bai, Haokun Lin, Jialin Zhao, Lu Hou, and Carlo Vittorio Cannistraci. Plug-and-play: An efficient post-training pruning method for large language models. In The Twelfth International Conference on Learning Representations, 2024 d . URL https://openreview.net/forum?id=Tr0lPx9woF

    2024