Pith. sign in

REVIEW 4 major objections 7 minor 47 references

Detecting and Pruning Prominent but Detrimental Neurons in Large Language Models

T0 review · 4 major / 7 minor · reviewed 2026-08-06 · deepseek-v4-flash

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

Pith's one-line read This paper claims that pruning the most highly attributed MLP neurons in a single layer, chosen using ten samples, improves LLM accuracy on multiple-choice benchmarks by removing dataset-specific shortcuts.

desk verdict Simple, cheap pruning trick with consistent gains across models and tasks, but the layer/percentage grid search on 10 samples needs a random-pruning control before the interpretation is fully solid. read the letter →

arxiv 2507.09185 v1 pith:DPX3HFM4 submitted 2025-07-12 cs.CL cs.LG

Ameen Ali , Shahar Katz , Lior Wolf , Ivan Titov This is my paper
classification cs.CLcs.LG
keywords neuronpruningIntegratedGradientsshortcutlearningdomain-specificmechanismsfew-shotadaptationlargelanguagemodelsmultiple-choicebenchmarksspuriouscorrelations
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

The paper tries to show that a small set of MLP neurons in transformer LLMs carry dataset-specific shortcuts: mechanisms that push the model's most confident logit but do not transfer. It proposes finding those neurons with Integrated Gradients computed against the maximum logit, then zeroing the top-scoring neurons in one layer chosen by accuracy on a ten-sample validation set. The claim is that this pruning-based adaptation improves accuracy across multiple-choice benchmarks, sometimes more than adaptation methods that use hundreds of labeled examples. If true, it means shortcut behavior is localized enough in a single feed-forward layer to be removed by zeroing weights rather than by retraining.

What carries the argument

The machinery is a weight-space Integrated Gradients attribution. For each question $q$, the paper defines $M(q)$ as the maximum output logit and, for neuron $n_j$ with weight $\hat{w}_j$, computes $IG(n_j,q) = \hat{w}_j \int_0^1 \partial M(q;\alpha\hat{w}_j)/\partial w_j\, d\alpha$, interpolating the neuron's weight from zero to its actual value. Scores are averaged over the adaptation set to form an $L \times d$ matrix, and the method then performs a grid search over layers and pruning percentages (5% to 50%), pruning the top-attributed neurons in the gate_proj layer of the selected layer and zeroing them. The argument-carrying step is that maximum-logit attribution, not label-based loss, identifies neurons that drive confident yet non-transferable predictions.

What would settle it

Run the full grid search on a fresh multiple-choice task and record test accuracy for every layer and pruning percentage. If the configuration selected by ten validation samples is not at or near the best test configuration, or if a large fraction of configurations cluster at the same test accuracy, then the reported gains would be selection noise rather than evidence that specific detrimental neurons were pruned.

Watch

Extended reading notes

Core claim

The central claim is that a small set of MLP neurons in large language models act as domain-specific mechanisms: they drive the model's maximum logit using spurious correlations, so they look highly influential under Integrated Gradients but do not generalize. The paper's Algorithm 1 computes Integrated Gradients with respect to the maximum logit for every neuron in every layer's gate_proj matrix, averages the scores over a small sample set, then grid-searches over layers and pruning percentages from 5% to 50%, pruning the top-attributed neurons and evaluating on the same small validation set. Reported results show accuracy gains across XNLI, MMLU, SST2, SST5, BoolQ, Balanced COPA, and XCOPA on LLaMA2-7B-Chat, LLaMA3.1-8B-Instruct, Mistral-7B-Instruct-v0.3, and Qwen2.5-7B-Instruct, while random pruning of equal size gives no comparable gains. The intended reading is that zeroing the most prominent-but-detrimental neurons, rather than retraining weights, restores the model's use of transferable representations.

Load-bearing premise

The method assumes that the single layer and pruning percentage that scores best on ten labeled samples will also be the best choice on the full test distribution, even though roughly three hundred configurations are compared on those ten samples.

Editorial extensions

If this is right

  • On the paper's evidence, a few hundred zeroed weights in one MLP layer can substitute for full fine-tuning on multiple-choice tasks, with no gradient updates and no label signal for the attribution step.
  • The gains reproduce across four model families and six benchmarks, which would mean shortcut mechanisms are a recurring, localized property of instruction-tuned transformers rather than an artifact of a single architecture.
  • Because random pruning of the same number of neurons does not improve accuracy, the effect would depend on identifying the right neurons, supporting the view that individual units encode dataset-specific behavior.
  • The reported cost of roughly 37 to 47 seconds of attribution computation per task would make this a practical few-shot adaptation recipe in data-scarce settings.

Reading between the lines

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

  • Editorial inference: the same maximum-logit pruning recipe could be tested on generation tasks and open-ended question answering, where high-confidence logits are less clearly defined; a win there would widen the claim beyond multiple-choice formats.
  • Editorial inference: the grid search's selection step is where selection noise can enter; a stronger confirmation would be to report whether the chosen layer and pruning percentage sits on a plateau of good test configurations or is an isolated argmax point.
  • Editorial inference: if the effect is truly about removing shortcuts, combining this pruning with activation steering should compound, since pruning deletes a pathway while steering adjusts directions; the paper does not test this combination.
  • Editorial inference: comparing neurons selected on correctly predicted versus incorrectly predicted adaptation samples, as the appendix begins to do, could refine the definition of a domain-specific neuron when labels are available.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

4 major / 7 minor

Summary. The paper proposes a few-shot adaptation method for transformer LLMs: compute Integrated Gradients of the maximum output logit with respect to gate_proj neuron weights over roughly 10 random training samples, rank neurons by their average attribution, grid-search over layers and pruning percentages on a small validation set, and then prune the top-attributed neurons in the selected layer. The authors report consistent accuracy gains over baseline models and over random pruning across LLaMA2, LLaMA3.1, Mistral, and Qwen on six English benchmarks and eight XCOPA languages, and they interpret these gains as evidence that pruning removes dataset-specific shortcut mechanisms (DSM neurons) and forces the model to rely on more generalizable representations.

Significance. If the empirical claim holds, the method is a novel and unusually cheap adaptation technique: it requires no gradient updates, very few samples, roughly 40 seconds of attribution computation, and it identifies a concrete set of neurons whose removal transfers to test data. The paper's strengths are its breadth of models and tasks, the explicit algorithm, the honest computational-cost table, and the robustness checks in Appendices D and E. The DSM-neuron construct is also falsifiable through the pruning intervention. However, the headline claim rests on a selection procedure over a 10-sample validation set, and the current evidence does not yet rule out selection noise as an alternative explanation; this is the main barrier to acceptance.

major comments (4)
  1. [Algorithm 1, lines 8–9 and Appendix D, Figure 3] The central claim is supported by selecting (l_opt, p_opt) as the argmax over Eval(Prune(M,l,p), Dval), where Dval is described in Section 1 as the same set of up to 10 samples used for IG attribution. With L≈28–32 and P={5%,...,50%}, this is roughly 300 evaluations on 10 labels, so the argmax is highly susceptible to selection noise; the test gains in Tables 1–3 are then the gains of the best of many noisy configurations. The only transfer analysis, Appendix D, reports a single dataset (SST2, LLaMA2), and the selected layer reaches 0% error on the adaptation set, which is consistent with overfitting on the selection criterion. Because the random-pruning baseline in Table 2 is not given the same grid-search selection procedure, it does not control for this multiple-comparisons effect. I would need an analysis of the full distribution of selected configurations over many random 10-sample adaptation sets, including tie counts, for each task and model, plus a random-pruning control that performs the same argmax, before the causal interpretation is established.
  2. [Tables 1–2 and Section 5.1] No error bars or significance tests are reported on the main results. Appendix E.1 reports variance only for BoolQ with LLaMA3.1-8B-Instruct (Table 7), and the ±0.65% at N=10 is small there, but that does not establish that the selected configurations on MMLU, SST2, XCOPA, or on LLaMA2, Mistral, and Qwen are stable. Given the 10-sample selection step, standard deviations on the reported test accuracies across repeated adaptation-set draws should be reported for every table row, not only for one model-task pair.
  3. [Equation (4), Section 3.3] The weight-space IG definition is not fully specified. The notation ∂M(q; α ŵ_j)/∂w_j is ambiguous because the path scales w_j by α while the derivative is taken with respect to w_j; the paper does not state the number of interpolation steps m, the exact baseline (zero weight for all neurons in the layer?), or how gradients are backpropagated through the gate_proj computation when the weight is scaled. These details determine the ranking of neurons and should be reported in the implementation section; without them the method is not reproducible as written.
  4. [Section 2 and Algorithm 1] The paper repeatedly claims that the method works 'without relying on annotated labels' (e.g., Section 2 and the XCOPA discussion in Section 5.1), but Algorithm 1 uses Eval(Prune(M,l,p), Dval) to select the layer and percentage, and accuracy evaluation requires labels. The IG scoring itself is label-free, but the grid-search selection is not. This overstates the method's contribution and should be corrected or qualified in the text and claims.
minor comments (7)
  1. [Section 3.4 and Algorithm 1] The grid-search ranges are inconsistent: Section 3.4 says 'increments of 5%, ranging from 5% to 40%' while Algorithm 1 states P = {5%, ..., 50%}; Appendix E.3 reports 10%–40% with additional points. The exact candidate set should be fixed because it is central to the selection-noise analysis.
  2. [Figure 2 and Table 1] The SST2 baseline for LLaMA2-7B-CHAT is given as 88.63% in Table 1 but as 88.87% in the Figure 2 legend; these numbers should be reconciled.
  3. [Table 6 caption] The caption of Table 6 says 'The number of data used for identifying key elements and testing,' but the table lists the dimensions of the intervention layers; this caption should be corrected.
  4. [Section 3.4] There is a typo, 'Algorthim 1', in the text preceding Algorithm 1; the algorithm itself is clear.
  5. [Models paragraph, Section 4] Qwen2.5-7B-Instruct is cited as 'Jiang et al., 2023a', which is the Mistral reference; a correct Qwen reference should be supplied.
  6. [Conclusions] The citation 'Sun et al.' in the conclusions is missing a year and full bibliographic entry.
  7. [Appendix D, Figure 3] The x-axis labels layers 0 through 30, but the model has 32 layers; please clarify whether the indexing is zero-based and whether layers 31 and 32 are omitted from the plot.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the pruning configuration is selected on a 10-sample adaptation set, but reported accuracies are on separate test sets, so the central result is not equivalent to its inputs by construction.

full rationale

The claimed derivation chain is: (1) attribute gate_proj neurons with Integrated Gradients w.r.t. the maximum logit (Eq. 4), (2) rank neurons by average attribution over a small adaptation set (Eq. 5), (3) choose layer and pruning percentage by argmax of validation accuracy (Algorithm 1, line 8), and (4) report accuracy on held-out test sets. Step (3) is hyperparameter/model selection on a 10-sample set; it does not make the test-set evaluation equal to the selection objective. The test sets are external to the adaptation set, so the reported gains are contingent empirical outcomes, not forced by construction. The paper even studies this transfer explicitly in Appendix D ('does the configuration that performs best on the adaptation samples maintain its effectiveness when applied to entirely new examples?'). The only self-citation (Ali et al., 2024) is used in Appendix E.2 as additional motivation for gate_proj, but Table 9 independently compares gate/up/down projections, so the citation is not load-bearing. Separate concerns--the 10-sample grid search may overfit (about 300 configurations), and the claim of not using labeled data is overstated because line 8 uses Dval accuracy--are correctness/robustness issues, not circularity. No equation is defined in terms of the result it is said to predict, and no fitted parameter is renamed as a prediction.

Assumptions & free parameters 4 free parameters · 4 assumptions · 1 invented entities

The central result rests on hyperparameters chosen by validation accuracy on 10 samples (l_opt, p_opt, N, and the unreported IG step count m), assumptions about what Integrated Gradients over weights measures, and the post hoc construct of DSM neurons. No new physical entities are introduced. The layer and percentage sweeps are the main free parameters. The mechanistic claim that the pruned neurons are dataset-specific is an assumption, not a measured property.

free parameters (4)
  • Layer index l_opt = per task, e.g., layer 10 for SST2 in Appendix D
    Selected by argmax validation accuracy over all layers (Algorithm 1 line 8); high risk of overfitting with only 10 adaptation samples.
  • Pruning percentage p_opt = per task, in 5% steps from 5% to 40-50%
    Selected jointly with l_opt on the same 10-sample adaptation set; the peak differs by task (e.g., 35% for BoolQ, 25% for SST2 in Appendix E.3).
  • IG interpolation steps m = not reported
    The Riemann-sum approximation in Eqs. (3) and (4) depends on m; the paper never reports m, so attribution scores are not exactly reproducible.
  • Number of IG samples N = 10 (default)
    Chosen by ablation (Figure 2); performance is similar for 5 to 100 samples, so this is a tuned rather than derived quantity.
assumptions (4)
  • domain assumption IG attribution to gate_proj weights reflects causal influence on the maximum logit.
    Section 3.3 uses IG scores to rank neurons for pruning; no causal verification is provided.
  • ad hoc to paper The top-attributed neurons in at least one layer are detrimental dataset-specific mechanisms, not generally useful reasoning neurons.
    This is the central mechanistic claim; the paper infers it from accuracy gains after pruning, but never measures dataset-specificity directly.
  • domain assumption The accuracy of a 10-sample adaptation set is a reliable selector for layer and pruning percentage.
    Algorithm 1 line 8 chooses l_opt and p_opt by validation accuracy; Appendix D provides only one transfer analysis (SST2), and the chosen layer reaches 0% adaptation error, suggesting possible overfitting.
  • standard math Zero weights are a valid absence baseline for weight-space Integrated Gradients.
    IG axioms assume a meaningful baseline; zeroing all gate_proj weights is an arbitrary reference for the MLP function and is not motivated by model semantics.
invented entities (1)
  • DSM neurons (domain-specific mechanism neurons)
    purpose: To name the hypothesized subset of MLP neurons that drive high-confidence but non-transferable predictions.
    The paper postulates this class and operationalizes it as top-IG neurons in a layer chosen by validation accuracy; no out-of-paper handle or mechanistic verification is given.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Detecting and Pruning Prominent but Detrimental Neurons in Large Language Models." pith.science (2026). https://pith.science/paper/DPX3HFM4

@misc{pith2026250709185,
  author       = {Pith},
  title        = {Pith review of: Detecting and Pruning Prominent but Detrimental Neurons in Large Language Models},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/DPX3HFM4}},
  note         = {Machine review of arXiv:2507.09185}
}
read the original abstract

Large language models (LLMs) often develop learned mechanisms specialized to specific datasets, such as reliance on domain-specific correlations, which yield high-confidence predictions without generalizable reasoning. While beneficial in one setting, these dataset-specific mechanisms typically degrade performance when models encounter novel tasks or distributions. In this work, we introduce a fine-tuning approach designed to enhance generalization by identifying and pruning neurons associated with dataset-specific mechanisms in transformer-based LLMs. Our method employs Integrated Gradients to quantify each neuron's influence on high-confidence predictions, pinpointing those that disproportionately contribute to dataset-specific performance without supporting robust, transferable reasoning. Selectively pruning these neurons compels the model to depend on generalizable representations. Evaluated across multiple-choice benchmarks, our pruning-based fine-tuning significantly enhances performance, surpassing prior (non-pruning) adaptation methods.

Figures

Figures reproduced from arXiv: 2507.09185 by the authors.

Figure 1
Figure 1. Comparative analysis of performance improvements across three LLM archi [PITH_FULL_IMAGE:figures/full_fig_p007_1.png] view at source ↗
Figure 2
Figure 2. Ablation studies over the number of samples used within the IG framework to [PITH_FULL_IMAGE:figures/full_fig_p009_2.png] view at source ↗
Figure 3
Figure 3. Comparison of error rates between pruned and non-pruned models across different [PITH_FULL_IMAGE:figures/full_fig_p018_3.png] view at source ↗

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

47 extracted references · 23 canonical work pages

  1. [1]

    Mitigating copy bias in in-context learning through neuron pruning

    Ameen Ali, Lior Wolf, and Ivan Titov. Mitigating copy bias in in-context learning through neuron pruning. arXiv preprint arXiv:2410.01288, 2024

    arXiv 2024

  2. [2]

    Impact of adversarial training on robustness and generalizability of language models

    Enes Altinisik, Hassan Sajjad, Husrev Taha Sencar, Safa Messaoud, and Sanjay Chawla. Impact of adversarial training on robustness and generalizability of language models. arXiv preprint arXiv:2211.05523, 2022

    arXiv 2022

  3. [3]

    will you find these shortcuts? a protocol for evaluating the faithfulness of input salience methods for text classification

    Jasmijn Bastings, Sebastian Ebert, Polina Zablotskaia, Anders Sandholm, and Katja Filippova. will you find these shortcuts? a protocol for evaluating the faithfulness of input salience methods for text classification. In Yoav Goldberg, Zornitsa Kozareva, and Yue Zhang (eds.), Proceedings of the 2022 Conference on Empirical Methods in Natural Language Proc...

    work page doi:10.18653/v1/2022.emnlp-ma 2022

  4. [4]

    A large annotated corpus for learning natural language inference

    Samuel R Bowman, Gabor Angeli, Christopher Potts, and Christopher D Manning. A large annotated corpus for learning natural language inference. arXiv preprint arXiv:1508.05326, 2015

    arXiv 2015

  5. [5]

    Language models are few-shot learners

    Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, et al. Language models are few-shot learners. Advances in neural information processing systems, 33: 0 1877--1901, 2020

    1901

  6. [6]

    Boolq: Exploring the surprising difficulty of natural yes/no questions

    Christopher Clark, Kenton Lee, Ming-Wei Chang, Tom Kwiatkowski, Michael Collins, and Kristina Toutanova. Boolq: Exploring the surprising difficulty of natural yes/no questions. arXiv preprint arXiv:1905.10044, 2019

    arXiv 1905

  7. [7]

    Knowledge neurons in pretrained transformers

    Damai Dai, Li Dong, Yaru Hao, Zhifang Sui, Baobao Chang, and Furu Wei. Knowledge neurons in pretrained transformers. In Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), Dublin, Ireland, May 2022. Association for Computational Linguistics

    work page 2022

  8. [8]

    BERT : Pre-training of deep bidirectional transformers for language understanding

    Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. BERT : Pre-training of deep bidirectional transformers for language understanding. 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: Human Language Technologies, Vol...

    doi:10.18653/v1/ 2019

Show all 47 references
  1. [9]

    Towards interpreting and mitigating shortcut learning behavior of nlu models

    Mengnan Du, Varun Manjunatha, Rajiv Jain, Ruchi Deshpande, Franck Dernoncourt, Jiuxiang Gu, Tong Sun, and Xia Hu. Towards interpreting and mitigating shortcut learning behavior of nlu models. In Proceedings of the 2021 Conference of the North American Chapter of the Associatio...

    2021

  2. [10]

    Shortcut learning of large language models in natural language understanding

    Mengnan Du, Fengxiang He, Na Zou, Dacheng Tao, and Xia Hu. Shortcut learning of large language models in natural language understanding. Communications of the ACM, 67 0 (1): 0 110--120, 2023

    2023

  3. [11]

    Data augmentations for improved (large) language model generalization

    Amir Feder, Yoav Wald, Claudia Shi, Suchi Saria, and David Blei. Data augmentations for improved (large) language model generalization. Advances in Neural Information Processing Systems, 36: 0 70638--70653, 2023

    2023

  4. [12]

    Finding dataset shortcuts with grammar induction

    Dan Friedman, Alexander Wettig, and Danqi Chen. Finding dataset shortcuts with grammar induction. In Yoav Goldberg, Zornitsa Kozareva, and Yue Zhang (eds.), Proceedings of the 2022 Conference on Empirical Methods in Natural Language Processing, pp.\ 4345--4363, Abu Dhabi, Unit...

    2022 doi

  5. [13]

    The llama 3 herd of models

    Aaron Grattafiori, Abhimanyu Dubey, Abhinav Jauhri, Abhinav Pandey, Abhishek Kadian, Ahmad Al-Dahle, Aiesha Letman, Akhil Mathur, Alan Schelten, Alex Vaughan, et al. The llama 3 herd of models. arXiv preprint arXiv:2407.21783, 2024

    2024 arXiv

  6. [14]

    Universal neurons in gpt2 language models

    Wes Gurnee, Theo Horsley, Zifan Carl Guo, Tara Rezaei Kheirkhah, Qinyi Sun, Will Hathaway, Neel Nanda, and Dimitris Bertsimas. Universal neurons in gpt2 language models. arXiv preprint arXiv:2401.12181, 2024

    2024 arXiv

  7. [15]

    Learning both weights and connections for efficient neural networks

    Song Han, Jeff Pool, John Tran, and William Dally. Learning both weights and connections for efficient neural networks. In Advances in neural information processing systems, volume 28, 2015

    2015

  8. [16]

    Parameter-efficient fine-tuning for large models: A comprehensive survey

    Zeyu Han, Chao Gao, Jinyang Liu, Jeff Zhang, and Sai Qian Zhang. Parameter-efficient fine-tuning for large models: A comprehensive survey. Transactions on Machine Learning Research, 2024. ISSN 2835-8856

    2024

  9. [17]

    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. arXiv preprint arXiv:2009.03300, 2020

    2009 arXiv

  10. [18]

    A survey on measuring and mitigating reasoning shortcuts in machine reading comprehension

    Xanh Ho, Johannes Mario Meissner, Saku Sugawara, and Akiko Aizawa. A survey on measuring and mitigating reasoning shortcuts in machine reading comprehension. arXiv preprint arXiv:2209.01824, 2022

    2022 arXiv

  11. [19]

    Parameter-efficient transfer learning for nlp

    Neil Houlsby, Andrei Giurgiu, Stanislaw Jastrzebski, Bruna Morrone, Quentin De Laroussilhe, Andrea Gesmundo, Mona Attariyan, and Sylvain Gelly. Parameter-efficient transfer learning for nlp. In International conference on machine learning, pp.\ 2790--2799. PMLR, 2019

    2019

  12. [20]

    Lora: Low-rank adaptation of large language models

    Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, Weizhu Chen, et al. Lora: Low-rank adaptation of large language models. ICLR, 1 0 (2): 0 3, 2022

    2022

  13. [21]

    Regional differences in synaptogenesis in human cerebral cortex

    Peter R Huttenlocher and Anand S Dabholkar. Regional differences in synaptogenesis in human cerebral cortex. Journal of Comparative Neurology, 387 0 (2): 0 167--178, 1997

    1997

  14. [22]

    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 ...

    2023 arXiv

  15. [23]

    Llm-blender: Ensembling large language models with pairwise comparison and generative fusion

    Dongfu Jiang, Xiang Ren, and Bill Yuchen Lin. Llm-blender: Ensembling large language models with pairwise comparison and generative fusion. In Proceedings of the 61th Annual Meeting of the Association for Computational Linguistics (ACL 2023), 2023 b

    2023

  16. [24]

    Optimal brain damage

    Yann LeCun, John S Denker, and Sara A Solla. Optimal brain damage. In Advances in neural information processing systems, pp.\ 598--605, 1990

    1990

  17. [25]

    Guiding LLM to fool itself: Automatically manipulating machine reading comprehension shortcut triggers

    Mosh Levy, Shauli Ravfogel, and Yoav Goldberg. Guiding LLM to fool itself: Automatically manipulating machine reading comprehension shortcut triggers. In The 2023 Conference on Empirical Methods in Natural Language Processing, 2023

    2023

  18. [26]

    Inference-time intervention: Eliciting truthful answers from a language model

    Kenneth Li, Oam Patel, Fernanda Vi \'e gas, Hanspeter Pfister, and Martin Wattenberg. Inference-time intervention: Eliciting truthful answers from a language model. Advances in Neural Information Processing Systems, 36: 0 41451--41530, 2023 a

    2023

  19. [27]

    Inference-time intervention: Eliciting truthful answers from a language model

    Kenneth Li, Oam Patel, Fernanda Vi \'e gas, Hanspeter Pfister, and Martin Wattenberg. Inference-time intervention: Eliciting truthful answers from a language model. In Thirty-seventh Conference on Neural Information Processing Systems, 2023 b . URL https://openreview.net/forum...

    2023

  20. [28]

    Training language models to follow instructions with human feedback

    Long Ouyang, Jeffrey Wu, Xu Jiang, Diogo Almeida, Carroll Wainwright, Pamela Mishkin, Chong Zhang, Sandhini Agarwal, Katarina Slama, Alex Ray, et al. Training language models to follow instructions with human feedback. Advances in neural information processing systems, 35: 0 2...

    2022

  21. [29]

    Extraordinary neoteny of synaptic spines in the human prefrontal cortex

    Zdravko Petanjek, Milo s Juda s , Goran S imi \'c , et al. Extraordinary neoteny of synaptic spines in the human prefrontal cortex. Proceedings of the National Academy of Sciences, 108 0 (32): 0 13281--13286, 2011

    2011

  22. [30]

    Ponti, Goran Glava s , Olga Majewska, Qianchu Liu, Ivan Vuli' c , and Anna Korhonen

    Edoardo M. Ponti, Goran Glava s , Olga Majewska, Qianchu Liu, Ivan Vuli' c , and Anna Korhonen. XCOPA: A multilingual dataset for causal commonsense reasoning. arXiv preprint, 2020. URL https://ducdauge.github.io/files/xcopa.pdf

    2020

  23. [31]

    Steering llama 2 via contrastive activation addition

    Nina Rimsky, Nick Gabrieli, Julian Schulz, Meg Tong, Evan Hubinger, and Alexander Turner. Steering llama 2 via contrastive activation addition. In Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), Bangkok, Thailand...

    2024

  24. [32]

    Choice of plausible alternatives: An evaluation of commonsense causal reasoning

    Melissa Roemmele, Cosmin Adrian Bejan, and Andrew S Gordon. Choice of plausible alternatives: An evaluation of commonsense causal reasoning. In 2011 AAAI Spring Symposium Series, 2011. URL https://people.ict.usc.edu/ gordon/publications/AAAI-SPRING11A.PDF

    2011

  25. [33]

    Recursive deep models for semantic compositionality over a sentiment treebank

    Richard Socher, Alex Perelygin, Jean Wu, Jason Chuang, Christopher D Manning, Andrew Y Ng, and Christopher Potts. Recursive deep models for semantic compositionality over a sentiment treebank. In Proceedings of the 2013 conference on empirical methods in natural language proce...

    2013

  26. [34]

    Confidence regulation neurons in language models

    Alessandro Stolfo, Ben Wu, Wes Gurnee, Yonatan Belinkov, Xingyi Song, Mrinmaya Sachan, and Neel Nanda. Confidence regulation neurons in language models. Advances in Neural Information Processing Systems, 37: 0 125019--125049, 2024

    2024

  27. [35]

    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

  28. [36]

    Axiomatic attribution for deep networks

    Mukund Sundararajan, Ankur Taly, and Qiqi Yan. Axiomatic attribution for deep networks. In International conference on machine learning, pp.\ 3319--3328. PMLR, 2017

    2017

  29. [37]

    Large language models can be lazy learners: Analyze shortcuts in in-context learning

    Ruixiang Tang, Dehan Kong, Longtao Huang, and Hui Xue. Large language models can be lazy learners: Analyze shortcuts in in-context learning. In Anna Rogers, Jordan Boyd-Graber, and Naoaki Okazaki (eds.), Findings of the Association for Computational Linguistics: ACL 2023, pp.\...

    2023 doi

  30. [38]

    Language-specific neurons: The key to multilingual capabilities in large language models

    Tianyi Tang, Wenyang Luo, Haoyang Huang, Dongdong Zhang, Xiaolei Wang, Xin Zhao, Furu Wei, and Ji-Rong Wen. Language-specific neurons: The key to multilingual capabilities in large language models. arXiv preprint arXiv:2402.16438, 2024

    2024 arXiv

  31. [39]

    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

    2023 arXiv

  32. [40]

    Neurons in large language models: Dead, n-gram, positional

    Elena Voita, Javier Ferrando, and Christoforos Nalmpantis. Neurons in large language models: Dead, n-gram, positional. arXiv preprint arXiv:2309.04827, 2023

    2023 arXiv

  33. [41]

    Semantics-adaptive activation intervention for LLM s via dynamic steering vectors

    Weixuan Wang, JINGYUAN YANG, and Wei Peng. Semantics-adaptive activation intervention for LLM s via dynamic steering vectors. In The Thirteenth International Conference on Learning Representations, 2025. URL https://openreview.net/forum?id=8WQ7VTfPTl

    2025

  34. [42]

    Do LLM s overcome shortcut learning? an evaluation of shortcut challenges in large language models

    Yu Yuan, Lili Zhao, Kai Zhang, Guangting Zheng, and Qi Liu. Do LLM s overcome shortcut learning? an evaluation of shortcut challenges in large language models. In Yaser Al-Onaizan and Yun-Nung Chen (eds.), Proceedings of the 2024 Conference on Empirical Methods in Natural Lang...

    2024

  35. [43]

    Comi: Correct and mitigate shortcut learning behavior in deep neural networks

    Lili Zhao, Qi Liu, Linan Yue, Wei Chen, Liyi Chen, Ruijun Sun, and Chao Song. Comi: Correct and mitigate shortcut learning behavior in deep neural networks. In Proceedings of the 47th International ACM SIGIR Conference on Research and Development in Information Retrieval, pp.\...

    2024

  36. [44]

    write newline

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

  37. [45]

    @esa (Ref

    \@ifxundefined[1] #1\@undefined \@firstoftwo \@secondoftwo \@ifnum[1] #1 \@firstoftwo \@secondoftwo \@ifx[1] #1 \@firstoftwo \@secondoftwo [2] @ #1 \@temptokena #2 #1 @ \@temptokena \@ifclassloaded agu2001 natbib The agu2001 class already includes natbib coding, so you should ...

  38. [46]

    \@lbibitem[] @bibitem@first@sw\@secondoftwo \@lbibitem[#1]#2 \@extra@b@citeb \@ifundefined br@#2\@extra@b@citeb \@namedef br@#2 \@nameuse br@#2\@extra@b@citeb \@ifundefined b@#2\@extra@b@citeb @num @parse #2 @tmp #1 NAT@b@open@#2 NAT@b@shut@#2 \@ifnum @merge>\@ne @bibitem@firs...

  39. [47]

    @open @close @open @close and [1] URL: #1 \@ifundefined chapter * \@mkboth \@ifxundefined @sectionbib * \@mkboth * \@mkboth\@gobbletwo \@ifclassloaded amsart * \@ifclassloaded amsbook * \@ifxundefined @heading @heading NAT@ctr thebibliography [1] @ \@biblabel @NAT@ctr \@bibset...

Pith tools

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