pith. sign in

arxiv: 2605.03052 · v2 · pith:OUDWHWEUnew · submitted 2026-05-04 · 💻 cs.CL

How Language Models Process Negation

Zhejian Zhou , Tianyi Zhou , Robin Jia , Jonathan May This is my paper

Pith reviewed 2026-07-01 00:07 UTC · model grok-4.3

classification 💻 cs.CL
keywords negation processinglanguage modelsmechanistic interpretabilityattention mechanismslarge language modelscausal analysisshortcut behavior
0
0 comments X

The pith

Language models process negation through both suppression of related concepts and construction of negative phrase representations, with construction dominant.

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

The paper establishes that LLMs contain internal components capable of correctly handling negation even when their final answers are wrong. Poor performance stems from late-layer attention that favors simple shortcuts, and removing those modules raises accuracy on negation questions. Models use two distinct mechanisms: attention heads that attend to a negated phrase and suppress associated ideas, and direct construction of a full negative representation such as turning "not gas" into a vector favoring liquids and solids. The construction route is shown to be more prominent through a combination of observational and causal tests on two open models.

Core claim

Even though open-weight models often give incorrect answers on negation questions, they possess internal components that process negation correctly. Their errors arise from late-layer attention behavior that promotes shortcuts; ablating those modules improves accuracy. Models implement both an attention-based suppression mechanism and a constructive mechanism that builds representations of entire negative phrases, with the constructive mechanism being more prominent.

What carries the argument

The constructive mechanism, in which models directly build a representation of the full negative phrase rather than only suppressing related concepts.

If this is right

  • Ablating late-layer attention modules substantially raises accuracy on questions involving negation.
  • Both suppression and construction mechanisms operate together inside the same models.
  • The constructive mechanism accounts for a larger share of negation processing than suppression.
  • The same pattern appears in both Mistral-7B and Llama-3.1-8B.

Where Pith is reading between the lines

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

  • The dominance of construction may extend to how models handle other logical operators such as conjunction or disjunction.
  • Targeted interventions that strengthen constructed representations could reduce shortcut reliance across a wider set of reasoning tasks.
  • If construction is the stronger route, training methods that reward explicit negative-phrase vectors might produce more robust negation behavior.

Load-bearing premise

The observational and causal interpretability techniques accurately isolate the causal role of specific attention heads and representations without unintended side effects on other computations.

What would settle it

If ablating the identified late-layer attention heads fails to improve accuracy on negation questions or if activation patterns show no constructed negative representations.

Figures

Figures reproduced from arXiv: 2605.03052 by Jonathan May, Robin Jia, Tianyi Zhou, Zhejian Zhou.

Figure 1
Figure 1. Figure 1: Illustration of competing mechanisms for negation. In the negation mechanism, attention module A1 moves the represen￾tation of the negation token (“not”) to the position of the concept being negated (e.g., amphibian). Subsequently, A2, together with downstream MLPs, constructs and promotes a new negated rep￾resentation (e.g., mammal). In contrast, the shortcut mechanism bypasses explicit negation reasoning… view at source ↗
Figure 2
Figure 2. Figure 2: OLMo2 Positive and Negative Accuracies at various pre-training checkpoints. We observe that negative accuracy first plummets at early training steps, then rises again and stabilizes. ablates their functionalities. If switching off certain heads recovers expected behaviors on P−, then we conclude that those attention modules play a causally significant role in shortcut behavior. More specifically, we apply … view at source ↗
Figure 3
Figure 3. Figure 3: Visualization of PCA subspace. The residual stream hidden states are taken from Llama-3.1-8B at layer 11 after the attention module (11 mid). The hidden states of P+ and P− are colored as blue and red. Arrows indicate the direction from one hidden state of P+ to the corresponding hidden state of P−. It can be seen that positive and negative hidden states are approximately linearly separable by one directio… view at source ↗
Figure 4
Figure 4. Figure 4: Attention Sink results performed on Llama-3.1-8B. The x-axis is the center of the window. The y-axis is the negation accuracy over the whole dataset. The dotted line is the negation accuracy of the vanilla model. 1) Attention modules around layer 14 are causally important. 2) Shortcut attention heads around layer 17 are identified. 3) Later layers play little role in negation. Results As shown in [PITH_FU… view at source ↗
Figure 5
Figure 5. Figure 5: Normalized evidence count plotted against attention layer index. The normalized evidence count measures the percentage of samples in the dataset for which evidence is identified at a given layer. Results are on Llama-3.1-8B. The blue (red) line indicates the ratio of samples that LogitLens identifies Y¯ (Y ) related tokens as top promoted (demoted) tokens. We observe that evidence count peaks at causally i… view at source ↗
Figure 5
Figure 5. Figure 5: Normalized evidence count plotted against attention layer index. The normalized evidence count measures the percentage of samples in the dataset for which evidence is identified at a given layer. Results are on Llama-3.1-8B. The blue (red) line indicates the ratio of samples that LogitLens identifies Y¯ (Y ) related tokens as top promoted (demoted) tokens. We observe that evidence count peaks at causally i… view at source ↗
Figure 6
Figure 6. Figure 6: Path Patching and Attention Sink Ablation results on Llama-3.1-8B. X axis indicates the center layer that we ablate or patch. Y axis is negation accuracy. Both methods suggest that mid-layer attention modules are causally important for negation processing. "justification": "These tokens suggest non-natural, persistent materials, which are incompatible with biodegradability." ,→ ,→ ,→ } ] ``` If **no convin… view at source ↗
Figure 6
Figure 6. Figure 6: Path Patching and Attention Sink Ablation results on Llama-3.1-8B. X axis indicates the center layer that we ablate or patch. Y axis is negation accuracy. Both methods suggest that mid-layer attention modules are causally important for negation processing. Your task is to identify **evidence that the model represents the semantic concept "not B"**, rather than merely attending to unrelated or noisy tokens.… view at source ↗
Figure 7
Figure 7. Figure 7: Path Patching and Attention Sink Ablation results on Mistral-7B. X axis indicates the center layer that we ablate or patch. Y axis is negation accuracy. Both methods suggest that mid-layer attention modules are causally important for negation processing. C. Additional Analyses C.1. Sanity Check on Negation Sensitivity Is sensitivity defined in Section 4.1 just an artifact of random￾ness? Is our choice of o… view at source ↗
Figure 7
Figure 7. Figure 7: Path Patching and Attention Sink Ablation results on Mistral-7B. X axis indicates the center layer that we ablate or patch. Y axis is negation accuracy. Both methods suggest that mid-layer attention modules are causally important for negation processing. * a brief justification explaining why they ,→ indicate "not B" Example: ```json [ { "layer": 14, "tokens": [" commercial", " ,→ Remaining"], "justificati… view at source ↗
Figure 8
Figure 8. Figure 8: Cross-validated LDA model accuracy as a function of position in the residual stream. A higher accuracy indicates that we can decode “not” from the residual stream more reliably. As shown, “not” is moved to “Y” at early to middle layers. C.5. LLM Annotation Results for Mistral-7B-v0.1 In view at source ↗
Figure 8
Figure 8. Figure 8: Cross-validated LDA model accuracy as a function of position in the residual stream. A higher accuracy indicates that we can decode “not” from the residual stream more reliably. As shown, “not” is moved to “Y” at early to middle layers. C. On Attention Sink and Attention Knockout Geva et al. (2023) propose Attention Knockout as a method to ablate attention modules moving information from spe￾cific spans of… view at source ↗
Figure 9
Figure 9. Figure 9: Normalized evidence count plotted against attention layer index. Results are on mistralai/Mistral-7B-v0.1. The blue (red) line indicates the ratio of samples that LogitLens identifies Y¯ (Y ) related tokens as top promoted (demoted) tokens. We observe that evidence count for “not” follows the same trends as patching results in view at source ↗
Figure 9
Figure 9. Figure 9: Normalized evidence count plotted against attention layer index. Results are on mistralai/Mistral-7B-v0.1. The blue (red) line indicates the ratio of samples that LogitLens identifies Y¯ (Y ) related tokens as top promoted (demoted) tokens. We observe that evidence count for “not” follows the same trends as patching results in [PITH_FULL_IMAGE:figures/full_fig_p019_9.png] view at source ↗
Figure 10
Figure 10. Figure 10: Visualization of the PCA space at different model layers. The hidden states of P+ and P− are colored as blue and red. Arrows indicate the direction from one hidden state of P+ to the corresponding hidden state of P−. It can be seen that positive and negative hidden states are approximately linearly separable by one direction. 18 view at source ↗
Figure 10
Figure 10. Figure 10: Negative and positive accuracies as a function of the sink layer for Cumulative Attention Sink. For each model, we sweep the layer from which the sink is applied (0-indexed, x-axis) and report both negative accuracy and positive accuracy. Dotted horizontal lines mark the vanilla (no-sink) accuracies, and the dashed green vertical line marks the sink layer that achieves the maximum negative accuracy. Acros… view at source ↗
Figure 11
Figure 11. Figure 11: Visualization of the PCA space at different model layers. The hidden states of P+ and P− are colored as blue and red. Arrows indicate the direction from one hidden state of P+ to the corresponding hidden state of P−. It can be seen that positive and negative hidden states are approximately linearly separable by one direction. 21 [PITH_FULL_IMAGE:figures/full_fig_p021_11.png] view at source ↗
read the original abstract

We study how Large Language Models (LLMs) process negation mechanistically. First, we establish that even though open-weight models often provide wrong answers to questions involving negation, they do possess internal components that process negation correctly. Their poor accuracy is due to late-layer attention behavior that promotes simple shortcuts; ablating those attention modules greatly improves accuracy on negation-related questions. Second, we uncover how models process negation. We consider two hypotheses: models could use attention heads that attend to the phrase being negated and suppress related concepts, or they could directly construct a representation of the entire negative phrase (e.g., representing "not gas" as a vector that promotes liquids and solids). We apply a range of observational and causal interpretability techniques on Mistral-7B and Llama-3.1-8B to show that models implement both mechanisms, with the "constructive" mechanism being more prominent. Combined, our work deepens the understanding of LLMs' internals, highlighting construction-dominant computations and the coexistence of competing mechanisms within LLMs.

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

1 major / 0 minor

Summary. The paper studies how LLMs process negation mechanistically on Mistral-7B and Llama-3.1-8B. It claims that models possess internal components that handle negation correctly, but late-layer attention promotes shortcuts leading to errors; ablating those modules improves accuracy on negation questions. Two mechanisms are hypothesized and tested: suppressive attention heads that attend to and suppress negated concepts, versus constructive mechanisms that build representations of negated phrases (e.g., 'not gas' promoting non-gas concepts). Using observational and causal interpretability techniques, the authors conclude both mechanisms coexist, with the constructive one more prominent.

Significance. If the mechanistic attributions hold, the work would advance understanding of how LLMs implement logical operations like negation by demonstrating the coexistence of competing internal mechanisms and the dominance of constructive computations. This could inform targeted interventions to improve logical consistency in models.

major comments (1)
  1. [Ablation and activation analysis sections] The central claim that models implement both suppressive and constructive negation mechanisms (with constructive dominant) rests on attention ablation and activation analyses to attribute causal roles to specific heads and representations. These methods require explicit controls or side-effect analyses to confirm they isolate the hypothesized circuits rather than producing indirect changes in unrelated components; without such evidence the comparative prominence conclusion is not securely supported. (Abstract; sections describing the causal interventions and results)

Simulated Author's Rebuttal

1 responses · 0 unresolved

We thank the referee for the constructive feedback on strengthening the causal evidence. We address the concern about controls for the ablation and activation analyses below.

read point-by-point responses

  1. Referee: [Ablation and activation analysis sections] The central claim that models implement both suppressive and constructive negation mechanisms (with constructive dominant) rests on attention ablation and activation analyses to attribute causal roles to specific heads and representations. These methods require explicit controls or side-effect analyses to confirm they isolate the hypothesized circuits rather than producing indirect changes in unrelated components; without such evidence the comparative prominence conclusion is not securely supported. (Abstract; sections describing the causal interventions and results)

    Authors: We agree that ablation and patching methods benefit from explicit controls to rule out indirect effects. Our study combined multiple converging methods: attention visualization and activation similarity for observational evidence, plus targeted ablation and activation patching for causal tests. We did compare negation-head ablations against random-head ablations of matched magnitude, finding that only the hypothesized heads produced large, task-specific gains on negation questions while random ablations had minimal effect. That said, we did not include exhaustive side-effect measurements on unrelated tasks or benchmarks. We will add these controls in revision, reporting performance changes on non-negation tasks after ablating the identified heads versus random controls, and will qualify the comparative prominence claim accordingly if the new results warrant it. revision: yes

Circularity Check

0 steps flagged

No significant circularity

full rationale

The paper's central claims rest on direct empirical experiments (attention ablation, activation analysis) applied to public models (Mistral-7B, Llama-3.1-8B). No mathematical derivation chain, self-defined quantities, fitted parameters renamed as predictions, or load-bearing self-citations appear in the provided text. The findings are presented as observational results from standard interpretability methods rather than any reduction to inputs by construction.

Axiom & Free-Parameter Ledger

0 free parameters · 2 axioms · 0 invented entities

The paper is empirical and relies on standard assumptions from the mechanistic interpretability literature rather than new free parameters or invented entities.

axioms (2)
  • domain assumption Ablation of specific attention modules isolates their causal contribution to output behavior without destroying unrelated capabilities
    Invoked when claiming that ablating late-layer attention greatly improves accuracy on negation questions.
  • domain assumption Activation and attention patterns observed via standard interpretability tools reflect the mechanisms used during normal forward passes
    Underlies both the identification of suppression vs. constructive mechanisms and the claim that internal components process negation correctly.

pith-pipeline@v0.9.1-grok · 5704 in / 1253 out tokens · 25973 ms · 2026-07-01T00:07:28.498963+00:00 · methodology

Review history (2 revisions) →

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

46 extracted references · 29 canonical work pages · 9 internal anchors

  1. [1]

    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 global.max substring 't := if while FUNCTION format.date year duplicate empty "emp...

  2. [2]

    Proceedings of the 2015 Conference on Empirical Methods in Natural Language Processing , month=

    Bowman, S. R., Angeli, G., Potts, C., and Manning, C. D. A large annotated corpus for learning natural language inference. In M \`a rquez, L., Callison-Burch, C., and Su, J. (eds.), Proceedings of the 2015 Conference on Empirical Methods in Natural Language Processing, pp.\ 632--642, Lisbon, Portugal, September 2015. Association for Computational Linguist...

    work page doi:10.18653/v1/d15-1075 2015

  3. [3]

    E., Hume, T., Carter, S., Henighan, T., and Olah, C

    Bricken, T., Templeton, A., Batson, J., Chen, B., Jermyn, A., Conerly, T., Turner, N., Anil, C., Denison, C., Askell, A., Lasenby, R., Wu, Y., Kravec, S., Schiefer, N., Maxwell, T., Joseph, N., Hatfield-Dodds, Z., Tamkin, A., Nguyen, K., McLean, B., Burke, J. E., Hume, T., Carter, S., Henighan, T., and Olah, C. Towards monosemanticity: Decomposing languag...

    2023

  4. [4]

    Summing up the facts: Additive mechanisms behind factual recall in llms, 2024

    Chughtai, B., Cooney, A., and Nanda, N. Summing up the facts: Additive mechanisms behind factual recall in llms, 2024. URL https://arxiv.org/abs/2402.07321

    work page arXiv 2024

  5. [5]

    Sparse Autoencoders Find Highly Interpretable Features in Language Models

    Cunningham, H., Ewart, A., Riggs, L., Huben, R., and Sharkey, L. Sparse autoencoders find highly interpretable features in language models, 2023. URL https://arxiv.org/abs/2309.08600

    work page internal anchor Pith review Pith/arXiv arXiv 2023

  6. [6]

    BERT : Pre-training of Deep Bidirectional Transformers for Language Understanding

    Devlin, J., Chang, M.-W., Lee, K., and Toutanova, K. BERT : Pre-training of deep bidirectional transformers for language understanding. In Burstein, J., Doran, C., and Solorio, T. (eds.), Proceedings of the 2019 Conference of the North A merican Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 (Long and Short...

    work page doi:10.18653/v1/n19-1423 2019

  7. [7]

    and Geva, M

    Elhelo, A. and Geva, M. Inferring functionality of attention heads from their parameters. In Che, W., Nabende, J., Shutova, E., and Pilehvar, M. T. (eds.), Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pp.\ 17701--17733, Vienna, Austria, July 2025. Association for Computational Linguistics...

    work page doi:10.18653/v1/2025.acl-long.866 2025

  8. [8]

    J., Liao, I., Gurnee, W., and Tegmark, M

    Engels, J., Michaud, E. J., Liao, I., Gurnee, W., and Tegmark, M. Not all language model features are one-dimensionally linear. In The Thirteenth International Conference on Learning Representations, 2025. URL https://openreview.net/forum?id=d63a4AM4hb

    2025

  9. [9]

    Fisher, R. A. The use of multiple measurements in taxonomic problems. Annals of Eugenics, 7 0 (2): 0 179--188, 1936. doi:https://doi.org/10.1111/j.1469-1809.1936.tb02137.x. URL https://onlinelibrary.wiley.com/doi/abs/10.1111/j.1469-1809.1936.tb02137.x

    work page doi:10.1111/j.1469-1809.1936.tb02137.x 1936

  10. [10]

    Transformer Feed-Forward Layers Are Key-Value Memories

    Geva, M., Schuster, R., Berant, J., and Levy, O. Transformer feed-forward layers are key-value memories. In Moens, M.-F., Huang, X., Specia, L., and Yih, S. W.-t. (eds.), Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing, pp.\ 5484--5495, Online and Punta Cana, Dominican Republic, November 2021. Association for Computa...

    work page internal anchor Pith review doi:10.18653/v1/2021.emnlp-main.446 2021

  11. [11]

    Dissecting Recall of Factual Associations in Auto-Regressive Language Models

    Geva, M., Bastings, J., Filippova, K., and Globerson, A. Dissecting recall of factual associations in auto-regressive language models. In Bouamor, H., Pino, J., and Bali, K. (eds.), Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing, pp.\ 12216--12235, Singapore, December 2023. Association for Computational Linguistics....

    work page doi:10.18653/v1/2023.emnlp-main.751 2023

  12. [12]

    The Llama 3 Herd of Models

    Grattafiori, A., Dubey, A., Jauhri, A., Pandey, A., Kadian, A., Al-Dahle, A., Letman, A., Mathur, A., Schelten, A., Vaughan, A., Yang, A., Fan, A., Goyal, A., Hartshorn, A., Yang, A., Mitra, A., Sravankumar, A., et al. The llama 3 herd of models, 2024. URL https://arxiv.org/abs/2407.21783

    work page internal anchor Pith review Pith/arXiv arXiv 2024

  13. [13]

    The unreasonable ineffectiveness of the deeper layers

    Gromov, A., Tirumala, K., Shapourian, H., Glorioso, P., and Roberts, D. The unreasonable ineffectiveness of the deeper layers. In The Thirteenth International Conference on Learning Representations, 2025. URL https://openreview.net/forum?id=ngmEcEer8a

    2025

  14. [14]

    and Handschuh, S

    Gubelmann, R. and Handschuh, S. Context matters: A pragmatic study of PLM s' negation understanding. In Muresan, S., Nakov, P., and Villavicencio, A. (eds.), Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pp.\ 4602--4621, Dublin, Ireland, May 2022. Association for Computational Linguistics....

    work page doi:10.18653/v1/2022.acl-long.315 2022

  15. [15]

    R., and Smith, N

    Gururangan, S., Swayamdipta, S., Levy, O., Schwartz, R., Bowman, S. R., and Smith, N. A. Annotation artifacts in natural language inference data. In Walker, M., Ji, H., and Stent, A. (eds.), Proceedings of the 2018 Conference of the North A merican Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 2 (Short Paper...

    work page doi:10.18653/v1/n18-2017 2018

  16. [16]

    Overthinking the truth: Understanding how language models process false demonstrations

    Halawi, D., Denain, J.-S., and Steinhardt, J. Overthinking the truth: Understanding how language models process false demonstrations. In Kim, B., Yue, Y., Chaudhuri, S., Fragkiadaki, K., Khan, M., and Sun, Y. (eds.), International Conference on Learning Representations, volume 2024, pp.\ 42749--42787, 2024. URL https://proceedings.iclr.cc/paper_files/pape...

    2024

  17. [17]

    and Glucksberg, S

    Hasson, U. and Glucksberg, S. Does understanding negation entail affirmation?: An examination of negated metaphors. Journal of Pragmatics, 38 0 (7): 0 1015--1032, 2006. ISSN 0378-2166. doi:https://doi.org/10.1016/j.pragma.2005.12.005. URL https://www.sciencedirect.com/science/article/pii/S0378216606000051. Special Issue: Processes and Products of Negation

    work page doi:10.1016/j.pragma.2005.12.005 2006

  18. [18]

    arXiv:2410.20526 (2024)

    He, Z., Shu, W., Ge, X., Chen, L., Wang, J., Zhou, Y., Liu, F., Guo, Q., Huang, X., Wu, Z., Jiang, Y.-G., and Qiu, X. Llama scope: Extracting millions of features from llama-3.1-8b with sparse autoencoders, 2024. URL https://arxiv.org/abs/2410.20526

    work page arXiv 2024

  19. [19]

    Hermann, K., Mobahi, H., FEL, T., and Mozer, M. C. On the foundations of shortcut learning. In The Twelfth International Conference on Learning Representations, 2024. URL https://openreview.net/forum?id=Tj3xLVuE9f

    2024

  20. [20]

    R., Eberle, O., Khakzar, A., and Nanda, N

    Jafari, F. R., Eberle, O., Khakzar, A., and Nanda, N. Relp: Faithful and efficient circuit discovery via relevance patching. In Mechanistic Interpretability Workshop at NeurIPS 2025, 2025. URL https://openreview.net/forum?id=5PKPy82sWN

    2025

  21. [21]

    Mistral 7B

    Jiang, A. Q., Sablayrolles, A., Mensch, A., Bamford, C., Chaplot, D. S., de las Casas, D., Bressand, F., Lengyel, G., Lample, G., Saulnier, L., Lavaud, L. R., Lachaux, M.-A., Stock, P., Scao, T. L., Lavril, T., Wang, T., Lacroix, T., and Sayed, W. E. Mistral 7b, 2023. URL https://arxiv.org/abs/2310.06825

    work page internal anchor Pith review Pith/arXiv arXiv 2023

  22. [22]

    and Sch \"u tze, H

    Kassner, N. and Sch \"u tze, H. Negated and misprimed probes for pretrained language models: Birds can talk, but cannot fly. In Jurafsky, D., Chai, J., Schluter, N., and Tetreault, J. (eds.), Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pp.\ 7811--7818, Online, July 2020. Association for Computational Linguistic...

    work page doi:10.18653/v1/2020.acl-main.698 2020

  23. [23]

    The self-contained negation test set

    Kletz, D., Amsili, P., and Candito, M. The self-contained negation test set. In Belinkov, Y., Hao, S., Jumelet, J., Kim, N., McCarthy, A., and Mohebbi, H. (eds.), Proceedings of the 6th BlackboxNLP Workshop: Analyzing and Interpreting Neural Networks for NLP, pp.\ 212--221, Singapore, December 2023. Association for Computational Linguistics. doi:10.18653/...

    work page doi:10.18653/v1/2023.blackboxnlp-1.16 2023

  24. [24]

    RoBERTa: A Robustly Optimized BERT Pretraining Approach

    Liu, Y., Ott, M., Goyal, N., Du, J., Joshi, M., Chen, D., Levy, O., Lewis, M., Zettlemoyer, L., and Stoyanov, V. Roberta: A robustly optimized bert pretraining approach, 2019. URL https://arxiv.org/abs/1907.11692

    work page internal anchor Pith review Pith/arXiv arXiv 2019

  25. [25]

    Don't think of the white bear: Ironic negation in transformer models under cognitive load, 2025

    Mann, L., Saxena, N., Tandon, S., Sun, C., Toteja, S., and Zhu, K. Don't think of the white bear: Ironic negation in transformer models under cognitive load, 2025. URL https://arxiv.org/abs/2511.12381

    work page arXiv 2025

  26. [26]

    and Tegmark, M

    Marks, S. and Tegmark, M. The geometry of truth: Emergent linear structure in large language model representations of true/false datasets. In First Conference on Language Modeling, 2024. URL https://openreview.net/forum?id=aajyHYjjsk

    2024

  27. [27]

    S., Conmy, A., Rushing, C., McGrath, T., and Nanda, N

    McDougall, C. S., Conmy, A., Rushing, C., McGrath, T., and Nanda, N. Copy suppression: Comprehensively understanding a motif in language model attention heads. In Belinkov, Y., Kim, N., Jumelet, J., Mohebbi, H., Mueller, A., and Chen, H. (eds.), Proceedings of the 7th BlackboxNLP Workshop: Analyzing and Interpreting Neural Networks for NLP, pp.\ 337--363,...

    work page doi:10.18653/v1/2024.blackboxnlp-1.22 2024

  28. [28]

    Locating and editing factual associations in gpt

    Meng, K., Bau, D., Andonian, A., and Belinkov, Y. Locating and editing factual associations in gpt. In Koyejo, S., Mohamed, S., Agarwal, A., Belgrave, D., Cho, K., and Oh, A. (eds.), Advances in Neural Information Processing Systems, volume 35, pp.\ 17359--17372. Curran Associates, Inc., 2022. URL https://proceedings.neurips.cc/paper_files/paper/2022/file...

    2022

  29. [29]

    interpreting gpt: the logit lens, 2020

    nostalgebraist. interpreting gpt: the logit lens, 2020. URL https://www.lesswrong.com/posts/AcKRB8wDpdaN6v6ru/interpreting-gpt-the-logit-lens

    2020

  30. [30]

    The building blocks of interpretability

    Olah, C., Satyanarayan, A., Johnson, I., Carter, S., Schubert, L., Ye, K., and Mordvintsev, A. The building blocks of interpretability. Distill, 2018. doi:10.23915/distill.00010. https://distill.pub/2018/building-blocks

    work page doi:10.23915/distill.00010 2018

  31. [31]

    The default computation of negated meanings

    Papeo, L., Hochmann, J.-R., and Battelli, L. The default computation of negated meanings. Journal of Cognitive Neuroscience, 28 0 (12): 0 1980--1986, 12 2016. ISSN 0898-929X. doi:10.1162/jocn_a_01016. URL https://doi.org/10.1162/jocn_a_01016

    work page doi:10.1162/jocn_a_01016 1980

  32. [32]

    Hypothesis only baselines in natural language inference

    Poliak, A., Naradowsky, J., Haldar, A., Rudinger, R., and Van Durme, B. Hypothesis only baselines in natural language inference. In Nissim, M., Berant, J., and Lenci, A. (eds.), Proceedings of the Seventh Joint Conference on Lexical and Computational Semantics, pp.\ 180--191, New Orleans, Louisiana, June 2018. Association for Computational Linguistics. do...

    work page doi:10.18653/v1/s18-2023 2018

  33. [33]

    Steering Llama 2 via Contrastive Activation Addition

    Rimsky, N., Gabrieli, N., Schulz, J., Tong, M., Hubinger, E., and Turner, A. Steering llama 2 via contrastive activation addition. In Ku, L.-W., Martins, A., and Srikumar, V. (eds.), Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pp.\ 15504--15522, Bangkok, Thailand, August 2024. Associatio...

    work page doi:10.18653/v1/2024.acl-long.828 2024

  34. [34]

    Gemma 2: Improving Open Language Models at a Practical Size

    Riviere, M., Pathak, S., Sessa, P. G., Hardin, C., Bhupatiraju, S., Hussenot, L., Mesnard, T., Shahriari, B., Ramé, A., Ferret, J., Liu, P., Tafti, P., Friesen, A., Casbon, M., Ramos, S., Kumar, R., Lan, C. L., Jerome, S., Tsitsulin, A., Vieillard, N., Stanczyk, P., Girgin, S., Momchev, N., Hoffman, M., Thakoor, S., Grill, J.-B., Neyshabur, B., Bachem, O....

    work page internal anchor Pith review Pith/arXiv arXiv 2024

  35. [35]

    R., Zhao, D., Patel, N

    Skean, O., Arefin, M. R., Zhao, D., Patel, N. N., Naghiyev, J., Lecun, Y., and Shwartz-Ziv, R. Layer by layer: Uncovering hidden representations in language models. In Singh, A., Fazel, M., Hsu, D., Lacoste-Julien, S., Berkenkamp, F., Maharaj, T., Wagstaff, K., and Zhu, J. (eds.), Proceedings of the 42nd International Conference on Machine Learning, volum...

    2025

  36. [36]

    S., Mueller, A., Wallace, B

    Todd, E., Li, M., Sharma, A. S., Mueller, A., Wallace, B. C., and Bau, D. Function vectors in large language models. In The Twelfth International Conference on Learning Representations, 2024. URL https://openreview.net/forum?id=AwyxtyMwaG

    2024

  37. [37]

    N., Kaiser, L

    Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones, L., Gomez, A. N., Kaiser, L. u., and Polosukhin, I. Attention is all you need. In Guyon, I., Luxburg, U. V., Bengio, S., Wallach, H., Fergus, R., Vishwanathan, S., and Garnett, R. (eds.), Advances in Neural Information Processing Systems, volume 30. Curran Associates, Inc., 2017. URL https://proc...

    2017

  38. [38]

    Walsh, E. P., Soldaini, L., Groeneveld, D., Lo, K., Arora, S., Bhagia, A., Gu, Y., Huang, S., Jordan, M., Lambert, N., Schwenk, D., Tafjord, O., Anderson, T., Atkinson, D., Brahman, F., Clark, C., Dasigi, P., Dziri, N., Ettinger, A., Guerquin, M., Heineman, D., Ivison, H., Koh, P. W., Liu, J., Malik, S., Merrill, W., Miranda, L. J. V., Morrison, J., Murra...

    2025

  39. [39]

    R., Variengien, A., Conmy, A., Shlegeris, B., and Steinhardt, J

    Wang, K. R., Variengien, A., Conmy, A., Shlegeris, B., and Steinhardt, J. Interpretability in the Wild : a Circuit for Indirect Object Identification in GPT -2 Small . In The Eleventh International Conference on Learning Representations , 2023. URL https://openreview.net/forum?id=NpsVSN6o4ul

    2023

  40. [40]

    Williams, A., Nangia, N., and Bowman, S. R. A broad-coverage challenge corpus for sentence understanding through inference. In Walker, M., Ji, H., and Stent, A. (eds.), Proceedings of the 2018 Conference of the North A merican Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 (Long Papers) , pp.\ 1112--1122, N...

    work page internal anchor Pith review doi:10.18653/v1/n18-1101 2018

  41. [41]

    Efficient streaming language models with attention sinks

    Xiao, G., Tian, Y., Chen, B., Han, S., and Lewis, M. Efficient streaming language models with attention sinks. In The Twelfth International Conference on Learning Representations, 2024. URL https://openreview.net/forum?id=NG7sS51zVF

    2024

  42. [42]

    Yan, T. L. and Jia, R. Promote, suppress, iterate: How language models answer one-to-many factual queries. In Christodoulopoulos, C., Chakraborty, T., Rose, C., and Peng, V. (eds.), Proceedings of the 2025 Conference on Empirical Methods in Natural Language Processing, pp.\ 16111--16134, Suzhou, China, November 2025. Association for Computational Linguist...

    work page doi:10.18653/v1/2025.emnlp-main.815 2025

  43. [43]

    Qwen2.5 Technical Report

    Yang, A., Yang, B., Zhang, B., Hui, B., Zheng, B., Yu, B., Li, C., Liu, D., Huang, F., Wei, H., Lin, H., Yang, J., Tu, J., Zhang, J., Yang, J., Yang, J., Zhou, J., Lin, J., Dang, K., Lu, K., Bao, K., Yang, K., Yu, L., Li, M., Xue, M., Zhang, P., Zhu, Q., Men, R., Lin, R., Li, T., Xia, T., Ren, X., Ren, X., Fan, Y., Su, Y., Zhang, Y., Wan, Y., Liu, Y., Cui...

    work page internal anchor Pith review Pith/arXiv arXiv 2024

  44. [44]

    Qwen3 Technical Report

    Yang, A., Li, A., Yang, B., Zhang, B., Hui, B., Zheng, B., Yu, B., Gao, C., Huang, C., Lv, C., Zheng, C., Liu, D., Zhou, F., Huang, F., Hu, F., Ge, H., Wei, H., Lin, H., Tang, J., Yang, J., Tu, J., Zhang, J., Yang, J., Yang, J., Zhou, J., Zhou, J., Lin, J., Dang, K., Bao, K., Yang, K., Yu, L., Deng, L., Li, M., Xue, M., Li, M., Zhang, P., Wang, P., Zhu, Q...

    work page internal anchor Pith review Pith/arXiv arXiv 2025

  45. [45]

    and Sennrich, R

    Zhang, B. and Sennrich, R. Root mean square layer normalization. In Wallach, H., Larochelle, H., Beygelzimer, A., d Alch\' e -Buc, F., Fox, E., and Garnett, R. (eds.), Advances in Neural Information Processing Systems, volume 32. Curran Associates, Inc., 2019. URL https://proceedings.neurips.cc/paper_files/paper/2019/file/1e8a19426224ca89e83cef47f1e7f53b-...

    2019

  46. [46]

    M., Gwilliams, L., King, J.-R., and Poeppel, D

    Zuanazzi, A., Ripollés, P., Lin, W. M., Gwilliams, L., King, J.-R., and Poeppel, D. Negation mitigates rather than inverts the neural representations of adjectives. PLOS Biology, 22 0 (5): 0 1--33, 05 2024. doi:10.1371/journal.pbio.3002622. URL https://doi.org/10.1371/journal.pbio.3002622

    work page doi:10.1371/journal.pbio.3002622 2024