---
title: "Embracing Anisotropy — paper by Roh et al. (2026)"
description: "They then propose a training-free rule that ranks dimensions by activation magnitude and keeps the top-k as Domain-Critical… 47 references."
authors: ["Youngji Roh", "Hyunjin Cho", "Jaehyung Kim"]
year: 2026
venue: "Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers)"
source: "https://aclanthology.org/2026.acl-long.1380/"
doi: "10.18653/v1/2026.acl-long.1380"
licence: "CC BY 4.0 (ACL Anthology)"
licence_url: "https://creativecommons.org/licenses/by/4.0/"
summary_licence: "Elamur Cosmic License"
canonical: "https://research.elamur.ai/free/en/papers/roh-2026-embracing-anisotropy-turning-massive-activations/"
---
[Elamur Research](https://md.research.elamur.ai/free/en/) · [this page in HTML](https://research.elamur.ai/free/en/papers/roh-2026-embracing-anisotropy-turning-massive-activations/)

[home](https://md.research.elamur.ai/free/en/) / [papers](https://md.research.elamur.ai/free/en/papers/) / [2026](https://md.research.elamur.ai/free/en/papers/2026/page/31/) / Embracing Anisotropy: Turning Massive Activations into Interpretable Control Knobs for Large Language Models

# Embracing Anisotropy: Turning Massive Activations into Interpretable Control Knobs for Large Language Models

**Authors:** Youngji Roh, Hyunjin Cho and Jaehyung Kim · **Year:** 2026 · **Venue:** [Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers)](https://md.research.elamur.ai/free/en/venues/acl/) ([ACL 2026](https://md.research.elamur.ai/free/en/venues/acl/2026/)) · **Publisher:** [Association for Computational Linguistics](https://md.research.elamur.ai/free/en/publishers/association-for-computational-linguistics/) · **Licence:** [CC BY 4.0 (ACL Anthology)](https://creativecommons.org/licenses/by/4.0/)

**Links:** [Publisher's page](https://aclanthology.org/2026.acl-long.1380/) · [PDF](https://aclanthology.org/2026.acl-long.1380.pdf) · [DOI 10.18653/v1/2026.acl-long.1380](https://doi.org/10.18653/v1/2026.acl-long.1380) · [OpenAlex](https://openalex.org/W7166809691)

**Citations:** 0 citations (OpenAlex, 2026-10-14)

## Summary
LLM hidden states are strongly anisotropic. A few feature dimensions carry "massive activations," far larger than all the others. Earlier work mostly treated these outliers as nuisances to suppress or preserve. This paper argues they are functional units that emerge from domain specialization.

On Gemma-2-2b-it and Qwen-3-8b, the authors use MMLU subjects as domains and make two observations. First, zeroing out a single dimension across all layers can sharply reduce accuracy. Second, some high-magnitude dimensions separate domains such as high school mathematics and biology. They then propose a training-free rule that ranks dimensions by activation magnitude and keeps the top-k as Domain-Critical Dimensions. These dimensions largely match the critical ones found by costly exhaustive masking.

Token-level inspection shows the dimensions act as readable detectors of mathematical symbols, biology vocabulary or topic words. Building on this, Critical Dimension Steering (CDS) applies activation steering only to those dimensions. CDS beats whole-dimension steering in both domain adaptation and jailbreaking. The work offers a cheap, interpretable control mechanism that needs no probes or sparse autoencoders.

## Key points
- Setup: all 57 MMLU subjects, with 100 test prompts per subject split evenly into an identification set and an evaluation set.
- Masking the single most harmful dimension cut Qwen-3-8b accuracy from 73.30% to 21.97% and Gemma-2-2b-it accuracy from 56.53% to 41.97%.
- Interpretable examples in Gemma-2-2b-it:
  - dim 1046 responds to math expressions;
  - dim 2106 responds to biology terms such as ATP and NAD;
  - dim 334 responds to subject keywords.
- CDS outperformed whole-dimension steering in 34 of 57 MMLU subjects. On AdvBench, it raised jailbreak attack success to 92%, compared with 84% for whole-dimension steering.
- Code is released on GitHub as "dimension-analyzer." The authors note limitations for tasks with fuzzy domain boundaries, such as creative writing.

*Our summary of the paper, in our own words — not its text. The work belongs to its authors; read it at the links above.*

## Institutions

- [Yonsei University](https://md.research.elamur.ai/free/en/universities/yonsei-university/) — University, Seoul, South Korea · 18 works · 501 citations

📚 Shelves (4): [🖥️ Computer Science](https://md.research.elamur.ai/free/en/shelves/computer-science/) · 18,614 works · [🤖 Artificial Intelligence](https://md.research.elamur.ai/free/en/shelves/computer-science/artificial-intelligence/) · 17,965 works · [🎭 Arts and Humanities](https://md.research.elamur.ai/free/en/shelves/arts-and-humanities/) · 17,450 works · [🗣️ Language and Linguistics](https://md.research.elamur.ai/free/en/shelves/arts-and-humanities/language-and-linguistics/) · 17,296 works

🧭 Topics (1): [🧠 Language models & pretraining](https://md.research.elamur.ai/free/en/topics/language-models/) · 3,894 works

## Cite

Cite as: Youngji Roh, Hyunjin Cho and Jaehyung Kim. 2026. Embracing Anisotropy: Turning Massive Activations into Interpretable Control Knobs for Large Language Models. Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers).

The paper itself, not this page. BibTeX from the library's record of the work.

```bibtex
@inproceedings{roh2026embracing,
  title = {{Embracing Anisotropy: Turning Massive Activations into Interpretable Control Knobs for Large Language Models}},
  author = {Youngji Roh, Hyunjin Cho and Jaehyung Kim},
  booktitle = {Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers)},
  year = {2026},
  doi = {10.18653/v1/2026.acl-long.1380},
  url = {https://doi.org/10.18653/v1/2026.acl-long.1380}
}
```

- [← previous Embedding Similarity Is Not Quality Estimation: Lessons from Replacing a Dedicated QE Model](https://md.research.elamur.ai/free/en/papers/zaikis-2026-embedding-similarity-is-not-quality/)
- [next → EmoAra: Emotion-Preserving English Speech Transcription and Cross-Lingual Translation with Arabic Text-to-Speech](https://md.research.elamur.ai/free/en/papers/hassan-2026-emoara-emotion-preserving-english-speech/)
- [↑ back to papers of 2026](https://md.research.elamur.ai/free/en/papers/2026/page/31/)

## References

47 references, 18 of them with a card of ours. ↗ leads to the work, «find» searches for one we have no id for yet.

1. Samira Abnar and Willem Zuidema. 2020. [Quantifying attention flow in transformers](https://md.research.elamur.ai/free/en/papers/abnar-2020-quantifying-attention-flow-transformers/). In Annual Meeting of the Association for Computational Linguistics (ACL).
2. Yongqi An, Xu Zhao, Tao Yu, Ming Tang, and Jinqiao Wang. 2025. Systematic outliers in large language models. In International Conference on Learning Representations (ICLR). [find](https://scholar.google.com/scholar?q=Yongqi+An%2C+Xu+Zhao%2C+Tao+Yu%2C+Ming+Tang%2C+and+Jinqiao+Wang%2E+2025%2E+Systematic+outliers+in+large+language+models%2E+In+International+Conference+on+Learning+Representations+%28ICLR%29%2E)
3. Trenton Bricken, Adly Templeton, Joshua Batson, Brian Chen, Adam Jermyn, Tom Conerly, Nicholas L Turner, Cem Anil, Carson Denison, Amanda Askell, Robert Lasenby, Yifan Wu, Shauna Kravec, Nicholas Schiefer, Tim Maxwell, Nicholas anthony Joseph, Alex Tamkin, Karina Nguyen, Brayden McLean, and 5 others… [find](https://scholar.google.com/scholar?q=Trenton+Bricken%2C+Adly+Templeton%2C+Joshua+Batson%2C+Brian+Chen%2C+Adam+Jermyn%2C+Tom+Conerly%2C+Nicholas+L+Turner%2C+Cem+Anil%2C+Carson+Denison%2C+Amanda+Askell%2C+Robert+Lasenby%2C+Yifan+Wu%2C+Shauna+Kravec%2C+Nicholas+Schi)
4. Collin Burns, Haotian Ye, Dan Klein, and Jacob Steinhardt. 2023. Discovering latent knowledge in language models without supervision. In International Conference on Learning Representations (ICLR). [find](https://scholar.google.com/scholar?q=Collin+Burns%2C+Haotian+Ye%2C+Dan+Klein%2C+and+Jacob+Steinhardt%2E+2023%2E+Discovering+latent+knowledge+in+language+models+without+supervision%2E+In+International+Conference+on+Learning+Representations+%28ICLR%29%2E)
5. Kevin Clark, Urvashi Khandelwal, Omer Levy, and Christopher D Manning. 2019. What does bert look at? an analysis of bert’s attention. In Proceedings of the 2019 ACL Workshop BlackboxNLP: Analyzing and Interpreting Neural Networks for NLP. [find](https://scholar.google.com/scholar?q=Kevin+Clark%2C+Urvashi+Khandelwal%2C+Omer+Levy%2C+and+Christopher+D+Manning%2E+2019%2E+What+does+bert+look+at%3F+an+analysis+of+bert%E2%80%99s+attention%2E+In+Proceedings+of+the+2019+ACL+Workshop+BlackboxNLP%3A+Analyzing+and)
6. Hoagy Cunningham, Aidan Ewart, Logan Riggs, Robert Huben, and Lee Sharkey. 2023. Sparse autoencoders find highly interpretable features in language models. In International Conference on Learning Representations (ICLR). [find](https://scholar.google.com/scholar?q=Hoagy+Cunningham%2C+Aidan+Ewart%2C+Logan+Riggs%2C+Robert+Huben%2C+and+Lee+Sharkey%2E+2023%2E+Sparse+autoencoders+find+highly+interpretable+features+in+language+models%2E+In+International+Conference+on+Learning+Repr)
7. Damai Dai, Li Dong, Yaru Hao, Zhifang Sui, Baobao Chang, and Furu Wei. 2022. Knowledge neurons in pretrained transformers. In Annual Meeting of the Association for Computational Linguistics (ACL), pages 8493–8502. [find](https://scholar.google.com/scholar?q=Damai+Dai%2C+Li+Dong%2C+Yaru+Hao%2C+Zhifang+Sui%2C+Baobao+Chang%2C+and+Furu+Wei%2E+2022%2E+Knowledge+neurons+in+pretrained+transformers%2E+In+Annual+Meeting+of+the+Association+for+Computational+Linguistics+%28ACL%29%2C+pag)
8. Tim Dettmers, Mike Lewis, Younes Belkada, and Luke Zettlemoyer. 2022. Llm. int8 (): 8-bit matrix multiplication for transformers at scale. In Advances in Neural Information Processing Systems (NeurIPS), volume 35, pages 30318–30332. [find](https://scholar.google.com/scholar?q=Tim+Dettmers%2C+Mike+Lewis%2C+Younes+Belkada%2C+and+Luke+Zettlemoyer%2E+2022%2E+Llm%2E+int8+%28%29%3A+8%2Dbit+matrix+multiplication+for+transformers+at+scale%2E+In+Advances+in+Neural+Information+Processing+Systems+%28NeurIPS)
9. Yue Ding, Karolis Martinkus, Damian Pascual, Simon Clematide, and Roger Wattenhofer. 2022. On isotropy calibration of transformer models. In Proceedings of the Third Workshop on Insights from Negative Results in NLP, pages 1–9. [find](https://scholar.google.com/scholar?q=Yue+Ding%2C+Karolis+Martinkus%2C+Damian+Pascual%2C+Simon+Clematide%2C+and+Roger+Wattenhofer%2E+2022%2E+On+isotropy+calibration+of+transformer+models%2E+In+Proceedings+of+the+Third+Workshop+on+Insights+from+Negative)
10. Nelson Elhage, Robert Lasenby, and Christopher Olah. 2023. Privileged bases in the transformer residual stream. Transformer Circuits Thread. [find](https://scholar.google.com/scholar?q=Nelson+Elhage%2C+Robert+Lasenby%2C+and+Christopher+Olah%2E+2023%2E+Privileged+bases+in+the+transformer+residual+stream%2E+Transformer+Circuits+Thread%2E)
11. Kawin Ethayarajh. 2019. How contextual are contextualized word representations? comparing the geometry of bert, elmo, and gpt-2 embeddings. In Conference on Empirical Methods in Natural Language Processing. [find](https://scholar.google.com/scholar?q=Kawin+Ethayarajh%2E+2019%2E+How+contextual+are+contextualized+word+representations%3F+comparing+the+geometry+of+bert%2C+elmo%2C+and+gpt%2D2+embeddings%2E+In+Conference+on+Empirical+Methods+in+Natural+Language+Proce)
12. Allyson Ettinger, Ahmed Elgohary, and Philip Resnik. 2016. [Probing for semantic evidence of composition by means of simple classification tasks](https://md.research.elamur.ai/free/en/papers/ettinger-2016-probing-semantic-evidence-composition-means/). In Proceedings of the 1st workshop on evaluating vector-space representations for nlp, pages 134–139.
13. Jun Gao, Di He, Xu Tan, Tao Qin, Liwei Wang, and Tie-Yan Liu. 2019. [Representation degeneration problem in training natural language generation models](https://md.research.elamur.ai/free/en/papers/gao-2019-representation-degeneration-problem-training-natural/). In International Conference on Learning Representations (ICLR).
14. 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, and 5 others.… [find](https://scholar.google.com/scholar?q=Leo+Gao%2C+Jonathan+Tow%2C+Baber+Abbasi%2C+Stella+Biderman%2C+Sid+Black%2C+Anthony+DiPofi%2C+Charles+Foster%2C+Laurence+Golding%2C+Jeffrey+Hsu%2C+Alain+Le+Noac%E2%80%99h%2C+Haonan+Li%2C+Kyle+McDonell%2C+Niklas+Muennighoff%2C+Chris+Oci)
15. Mor Geva, Roei Schuster, Jonathan Berant, and Omer Levy. 2021. Transformer feed-forward layers are key-value memories. In Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 5484–5495. [find](https://scholar.google.com/scholar?q=Mor+Geva%2C+Roei+Schuster%2C+Jonathan+Berant%2C+and+Omer+Levy%2E+2021%2E+Transformer+feed%2Dforward+layers+are+key%2Dvalue+memories%2E+In+Conference+on+Empirical+Methods+in+Natural+Language+Processing+%28EMNLP%29%2C+pages+)
16. Aaron Grattafiori, Abhimanyu Dubey, Abhinav Jauhri, Abhinav Pandey, Abhishek Kadian, Ahmad Al-Dahle, Aiesha Letman, Akhil Mathur, Alan Schelten, Alex Vaughan, and 1 others. 2024. [The llama 3 herd of models](https://md.research.elamur.ai/free/en/papers/grattafiori-2024-llama-3-herd-models/). arXiv preprint arXiv:2407.21783.
17. Katharina Hämmerl, Alina Fastowski, Jindřich Libovickỳ, and Alexander Fraser. 2023. Exploring anisotropy and outliers in multilingual language models for cross-lingual semantic sentence similarity. In Annual Meeting of the Association for Computational Linguistics (ACL). [find](https://scholar.google.com/scholar?q=Katharina+H%C3%A4mmerl%2C+Alina+Fastowski%2C+Jindr%CC%8Cich+Libovicky%CC%80%2C+and+Alexander+Fraser%2E+2023%2E+Exploring+anisotropy+and+outliers+in+multilingual+language+models+for+cross%2Dlingual+semantic+sentence+similarity%2E+)
18. Bobby He, Lorenzo Noci, Daniele Paliotta, Imanol Schlag, and Thomas Hofmann. 2024. Understanding and minimising outlier features in neural network training. In Advances in Neural Information Processing Systems (NeurIPS). [find](https://scholar.google.com/scholar?q=Bobby+He%2C+Lorenzo+Noci%2C+Daniele+Paliotta%2C+Imanol+Schlag%2C+and+Thomas+Hofmann%2E+2024%2E+Understanding+and+minimising+outlier+features+in+neural+network+training%2E+In+Advances+in+Neural+Information+Processin)
19. Dan Hendrycks, Collin Burns, Steven Basart, Andy Zou, Mantas Mazeika, Dawn Song, and Jacob Steinhardt. 2021. [Measuring massive multitask language understanding](https://md.research.elamur.ai/free/en/papers/hendrycks-2020-measuring-massive-multitask-language-understanding/). In International Conference on Learning Representations (ICLR).
20. Olga Kovaleva, Saurabh Kulshreshtha, Anna Rogers, and Anna Rumshisky. 2021. [Bert busters: Outlier dimensions that disrupt transformers](https://md.research.elamur.ai/free/en/papers/kovaleva-2021-bert-busters-outlier-dimensions-that/). In Findings of the Association for Computational Linguistics: ACL-IJCNLP 2021.
21. Bruce W Lee, Inkit Padhi, Karthikeyan Natesan Ramamurthy, Erik Miehling, Pierre Dognin, Manish Nagireddy, and Amit Dhurandhar. 2024. Programming refusal with conditional activation steering. In International Conference on Learning Representations (ICLR). [find](https://scholar.google.com/scholar?q=Bruce+W+Lee%2C+Inkit+Padhi%2C+Karthikeyan+Natesan+Ramamurthy%2C+Erik+Miehling%2C+Pierre+Dognin%2C+Manish+Nagireddy%2C+and+Amit+Dhurandhar%2E+2024%2E+Programming+refusal+with+conditional+activation+steering%2E+In+Intern)
22. Kenneth Li, Oam Patel, Fernanda Viégas, Hanspeter Pfister, and Martin Wattenberg. 2023. [Inference-time intervention: Eliciting truthful answers from a language model](https://md.research.elamur.ai/free/en/papers/li-2023-inference-time-intervention-eliciting-truthful/). In Advances in Neural Information Processing Systems (NeurIPS), volume 36, pages 41451–41530.
23. Kevin Meng, David Bau, Alex Andonian, and Yonatan Belinkov. 2022. Locating and editing factual associations in gpt. In Advances in Neural Information Processing Systems (NeurIPS), volume 35, pages 17359–17372. [find](https://scholar.google.com/scholar?q=Kevin+Meng%2C+David+Bau%2C+Alex+Andonian%2C+and+Yonatan+Belinkov%2E+2022%2E+Locating+and+editing+factual+associations+in+gpt%2E+In+Advances+in+Neural+Information+Processing+Systems+%28NeurIPS%29%2C+volume+35%2C+pages+173)
24. Team OLMo, Pete Walsh, Luca Soldaini, Dirk Groeneveld, Kyle Lo, Shane Arora, Akshita Bhagia, Yuling Gu, Shengyi Huang, Matt Jordan, and 1 others. 2024. 2 olmo 2 furious. arXiv preprint arXiv:2501.00656. [find](https://scholar.google.com/scholar?q=Team+OLMo%2C+Pete+Walsh%2C+Luca+Soldaini%2C+Dirk+Groeneveld%2C+Kyle+Lo%2C+Shane+Arora%2C+Akshita+Bhagia%2C+Yuling+Gu%2C+Shengyi+Huang%2C+Matt+Jordan%2C+and+1+others%2E+2024%2E+2+olmo+2+furious%2E+arXiv+preprint+arXiv%3A2501%2E0065)
25. Anton Razzhigaev, Matvey Mikhalchuk, Elizaveta Goncharova, Ivan Oseledets, Denis Dimitrov, and Andrey Kuznetsov. 2024. The shape of learning: Anisotropy and intrinsic dimensions in transformer-based models. In Conference of the European Chapter of the Association for Computational Linguistics (EACL)… [find](https://scholar.google.com/scholar?q=Anton+Razzhigaev%2C+Matvey+Mikhalchuk%2C+Elizaveta+Goncharova%2C+Ivan+Oseledets%2C+Denis+Dimitrov%2C+and+Andrey+Kuznetsov%2E+2024%2E+The+shape+of+learning%3A+Anisotropy+and+intrinsic+dimensions+in+transformer%2Dbased+m)
26. Nina Rimsky, Nick Gabrieli, Julian Schulz, Meg Tong, Evan Hubinger, and Alexander Turner. 2024. Steering llama 2 via contrastive activation addition. In Annual Meeting of the Association for Computational Linguistics (ACL), pages 15504–15522. [find](https://scholar.google.com/scholar?q=Nina+Rimsky%2C+Nick+Gabrieli%2C+Julian+Schulz%2C+Meg+Tong%2C+Evan+Hubinger%2C+and+Alexander+Turner%2E+2024%2E+Steering+llama+2+via+contrastive+activation+addition%2E+In+Annual+Meeting+of+the+Association+for+Computati)
27. William Rudman, Catherine Chen, and Carsten Eickhoff. 2023. Outlier dimensions encode task specific knowledge. In Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing, Singapore. Association for Computational Linguistics. [find](https://scholar.google.com/scholar?q=William+Rudman%2C+Catherine+Chen%2C+and+Carsten+Eickhoff%2E+2023%2E+Outlier+dimensions+encode+task+specific+knowledge%2E+In+Proceedings+of+the+2023+Conference+on+Empirical+Methods+in+Natural+Language+Processing)
28. William Rudman and Carsten Eickhoff. 2023. Stable anisotropic regularization. In International Conference on Learning Representations (ICLR). [find](https://scholar.google.com/scholar?q=William+Rudman+and+Carsten+Eickhoff%2E+2023%2E+Stable+anisotropic+regularization%2E+In+International+Conference+on+Learning+Representations+%28ICLR%29%2E)
29. Shivalika Singh, Angelika Romanou, Clémentine Fourrier, David Ifeoluwa Adelani, Jian Gang Ngui, Daniel Vila-Suero, Peerat Limkonchotiwat, Kelly Marchisio, Wei Qi Leong, Yosephine Susanto, and 1 others. 2025. [Global mmlu: Understanding and addressing cultural and linguistic biases in multilingual eva…](https://md.research.elamur.ai/free/en/papers/singh-2025-global-mmlu-understanding-addressing-cultural/)
30. Mingjie Sun, Xinlei Chen, J Zico Kolter, and Zhuang Liu. 2024. [Massive activations in large language models](https://md.research.elamur.ai/free/en/papers/sun-2024-massive-activations-large-language-models/). In Conference on Language Modeling (COLM).
31. Gemma Team, Morgane Riviere, Shreya Pathak, Pier Giuseppe Sessa, Cassidy Hardin, Surya Bhupatiraju, Léonard Hussenot, Thomas Mesnard, Bobak Shahriari, Alexandre Ramé, and 1 others. 2024. Gemma 2: Improving open language models at a practical size. arXiv preprint arXiv:2408.00118. [find](https://scholar.google.com/scholar?q=Gemma+Team%2C+Morgane+Riviere%2C+Shreya+Pathak%2C+Pier+Giuseppe+Sessa%2C+Cassidy+Hardin%2C+Surya+Bhupatiraju%2C+L%C3%A9onard+Hussenot%2C+Thomas+Mesnard%2C+Bobak+Shahriari%2C+Alexandre+Ram%C3%A9%2C+and+1+others%2E+2024%2E+Gemma+2%3A+Impr)
32. William Timkey and Marten van Schijndel. 2021. [All bark and no bite: Rogue dimensions in transformer language models obscure representational quality](https://md.research.elamur.ai/free/en/papers/timkey-2021-all-bark-no-bite-rogue/). In Conference on Empirical Methods in Natural Language Processing (EMNLP).
33. Alexander Matt Turner, Lisa Thiergart, Gavin Leech, David Udell, Juan J Vazquez, Ulisse Mini, and Monte MacDiarmid. 2023. [Steering language models with activation engineering](https://md.research.elamur.ai/free/en/papers/turner-2023-steering-language-models-activation-engineering/). arXiv preprint arXiv:2308.10248.
34. Jesse Vig, Sebastian Gehrmann, Yonatan Belinkov, Sharon Qian, Daniel Nevo, Yaron Singer, and Stuart Shieber. 2020. Investigating gender bias in language models using causal mediation analysis. In Advances in Neural Information Processing Systems (NeurIPS), volume 33, pages 12388–12401. [find](https://scholar.google.com/scholar?q=Jesse+Vig%2C+Sebastian+Gehrmann%2C+Yonatan+Belinkov%2C+Sharon+Qian%2C+Daniel+Nevo%2C+Yaron+Singer%2C+and+Stuart+Shieber%2E+2020%2E+Investigating+gender+bias+in+language+models+using+causal+mediation+analysis%2E+In+Adva)
35. Elena Voita, David Talbot, Fedor Moiseev, Rico Sennrich, and Ivan Titov. 2019. [Analyzing multi-head self-attention: Specialized heads do the heavy lifting, the rest can be pruned](https://md.research.elamur.ai/free/en/papers/voita-2019-analyzing-multi-head-self-attention/). In Annual Meeting of the Association for Computational Linguistics (ACL).
36. Alexander Wei, Nika Haghtalab, and Jacob Steinhardt. 2023a. [Jailbroken: How does llm safety training fail?](https://md.research.elamur.ai/free/en/papers/wei-2023-jailbroken-llm-safety-training-fail/) In Advances in Neural Information Processing Systems (NeurIPS).
37. Xiuying Wei, Yunchen Zhang, Yuhang Li, Xiangguo Zhang, Ruihao Gong, Jinyang Guo, and Xianglong Liu. 2023b. Outlier suppression+: Accurate quantization of large language models by equivalent and optimal shifting and scaling. In Conference on Empirical Methods in Natural Language Processing (EMNLP). [find](https://scholar.google.com/scholar?q=Xiuying+Wei%2C+Yunchen+Zhang%2C+Yuhang+Li%2C+Xiangguo+Zhang%2C+Ruihao+Gong%2C+Jinyang+Guo%2C+and+Xianglong+Liu%2E+2023b%2E+Outlier+suppression%2B%3A+Accurate+quantization+of+large+language+models+by+equivalent+and+optima)
38. Xiuying Wei, Yunchen Zhang, Xiangguo Zhang, Ruihao Gong, Shanghang Zhang, Qi Zhang, Fengwei Yu, and Xianglong Liu. 2022. Outlier suppression: Pushing the limit of low-bit transformer language models. In Advances in Neural Information Processing Systems (NeurIPS), volume 35, pages 17402–17414. [find](https://scholar.google.com/scholar?q=Xiuying+Wei%2C+Yunchen+Zhang%2C+Xiangguo+Zhang%2C+Ruihao+Gong%2C+Shanghang+Zhang%2C+Qi+Zhang%2C+Fengwei+Yu%2C+and+Xianglong+Liu%2E+2022%2E+Outlier+suppression%3A+Pushing+the+limit+of+low%2Dbit+transformer+language+models%2E+)
39. Guangxuan Xiao, Ji Lin, Mickael Seznec, Hao Wu, Julien Demouth, and Song Han. 2023. Smoothquant: Accurate and efficient post-training quantization for large language models. In International conference on machine learning, pages 38087–38099. PMLR. [find](https://scholar.google.com/scholar?q=Guangxuan+Xiao%2C+Ji+Lin%2C+Mickael+Seznec%2C+Hao+Wu%2C+Julien+Demouth%2C+and+Song+Han%2E+2023%2E+Smoothquant%3A+Accurate+and+efficient+post%2Dtraining+quantization+for+large+language+models%2E+In+International+conferenc)
40. An Yang, Anfeng Li, Baosong Yang, Beichen Zhang, Binyuan Hui, Bo Zheng, Bowen Yu, Chang Gao, Chengen Huang, Chenxu Lv, and 1 others. 2025. [Qwen3 technical report](https://md.research.elamur.ai/free/en/papers/yang-2025-qwen3-technical-report/). arXiv preprint arXiv:2505.09388.
41. Zhewei Yao, Reza Yazdani Aminabadi, Minjia Zhang, Xiaoxia Wu, Conglong Li, and Yuxiong He. 2022. [Zeroquant: Efficient and affordable post-training quantization for large-scale transformers](https://md.research.elamur.ai/free/en/papers/yao-2022-zeroquant-efficient-affordable-post-training/). In Advances in Neural Information Processing Systems (NeurIPS).
42. Zeyu Yun, Yubei Chen, Bruno Olshausen, and Yann Le-Cun. 2021. Transformer visualization via dictionary learning: contextualized embedding as a linear superposition of transformer factors. In Proceedings of Deep Learning Inside Out (DeeLIO): The 2nd Workshop on Knowledge Extraction and Integration fo… [find](https://scholar.google.com/scholar?q=Zeyu+Yun%2C+Yubei+Chen%2C+Bruno+Olshausen%2C+and+Yann+Le%2DCun%2E+2021%2E+Transformer+visualization+via+dictionary+learning%3A+contextualized+embedding+as+a+linear+superposition+of+transformer+factors%2E+In+Proceedin)
43. Zhihao Zhang, Jun Zhao, Qi Zhang, Tao Gui, and Xuan-Jing Huang. 2024. [Unveiling linguistic regions in large language models](https://md.research.elamur.ai/free/en/papers/zhang-2024-unveiling-linguistic-regions-large-language/). In Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 6228– 6247.
44. Tianyang Zhao, Kunwar Yashraj Singh, Srikar Appalaraju, Peng Tang, Ying Nian Wu, and Li Erran Li. 2025. On the analysis and distillation of emergent outlier properties in pre-trained language models. In Proceedings of the 2025 Conference of the Nations of the Americas Chapter of the Association for… [find](https://scholar.google.com/scholar?q=Tianyang+Zhao%2C+Kunwar+Yashraj+Singh%2C+Srikar+Appalaraju%2C+Peng+Tang%2C+Ying+Nian+Wu%2C+and+Li+Erran+Li%2E+2025%2E+On+the+analysis+and+distillation+of+emergent+outlier+properties+in+pre%2Dtrained+language+models%2E+)
45. Zhanxing Zhu, Jingfeng Wu, Bing Yu, Lei Wu, and Jinwen Ma. 2018. The anisotropic noise in stochastic gradient descent: Its behavior of escaping from sharp minima and regularization effects. In Proceedings of the International Conference on Machine Learning (ICML). [find](https://scholar.google.com/scholar?q=Zhanxing+Zhu%2C+Jingfeng+Wu%2C+Bing+Yu%2C+Lei+Wu%2C+and+Jinwen+Ma%2E+2018%2E+The+anisotropic+noise+in+stochastic+gradient+descent%3A+Its+behavior+of+escaping+from+sharp+minima+and+regularization+effects%2E+In+Proceed)
46. Andy Zou, Long Phan, Sarah Chen, James Campbell, Phillip Guo, Richard Ren, Alexander Pan, Xuwang Yin, Mantas Mazeika, Ann-Kathrin Dombrowski, and 1 others. 2023a. [Representation engineering: A top-down approach to ai transparency](https://md.research.elamur.ai/free/en/papers/zou-2023-representation-engineering-top-down-approach/). arXiv preprint arXiv:2310.01405.
47. Andy Zou, Zifan Wang, Nicholas Carlini, Milad Nasr, J Zico Kolter, and Matt Fredrikson. 2023b. [Universal and transferable adversarial attacks on aligned language models](https://md.research.elamur.ai/free/en/papers/zou-2023-universal-transferable-adversarial-attacks-aligned/). arXiv preprint arXiv:2307.15043.

*Updated 2 November 2026*