---
title: "The Unreasonable Ineffectiveness of the Deeper Layers (2024)"
description: "The paper asks how much of an open-weight LLM's depth is actually needed, and uses layer removal… 79 references, cited by 64 papers here."
authors: ["Andrey Gromov", "Kushal Tirumala", "Hassan Shapourian", "Paolo Glorioso", "Daniel A. Roberts"]
year: 2024
venue: "arXiv:2403.17887"
source: "https://arxiv.org/abs/2403.17887"
doi: "10.48550/arxiv.2403.17887"
arxiv: "2403.17887"
licence: "arXiv — non-exclusive license to distribute"
licence_url: "http://arxiv.org/licenses/nonexclusive-distrib/1.0/"
summary_licence: "Elamur Cosmic License"
canonical: "https://research.elamur.ai/free/en/papers/gromov-2024-unreasonable-ineffectiveness-deeper-layers/"
---
[Elamur Research](https://md.research.elamur.ai/free/en/) · [this page in HTML](https://research.elamur.ai/free/en/papers/gromov-2024-unreasonable-ineffectiveness-deeper-layers/)

[home](https://md.research.elamur.ai/free/en/) / [papers](https://md.research.elamur.ai/free/en/papers/) / [2024](https://md.research.elamur.ai/free/en/papers/2024/page/4/) / The Unreasonable Ineffectiveness of the Deeper Layers

# The Unreasonable Ineffectiveness of the Deeper Layers

**Authors:** Andrey Gromov, [Kushal Tirumala](https://md.research.elamur.ai/free/en/authors/kushal-tirumala/), Hassan Shapourian, Paolo Glorioso and Daniel A. Roberts · **Year:** 2024 · **Venue:** arXiv:2403.17887 · **Licence:** [arXiv — non-exclusive license to distribute](http://arxiv.org/licenses/nonexclusive-distrib/1.0/)

**Links:** [Publisher's page](https://arxiv.org/abs/2403.17887) · [PDF](https://arxiv.org/pdf/2403.17887) · [DOI 10.48550/arxiv.2403.17887](https://doi.org/10.48550/arxiv.2403.17887) · [arXiv 2403.17887](https://arxiv.org/abs/2403.17887) · [OpenAlex](https://openalex.org/W4393248014)

**Citations:** 5 citations (OpenAlex, 2026-10-14) · [64](https://md.research.elamur.ai/free/en/papers/gromov-2024-unreasonable-ineffectiveness-deeper-layers/cited-by/) in our library, of the 141,840 papers whose bibliographies we read

## Summary
The paper asks how much of an open-weight LLM's depth is actually needed, and uses layer removal as a probe of where knowledge lives. The authors measure the angular distance between the representation entering a block of n consecutive layers and the one leaving it, then drop the block whose input and output are most alike. A brief QLoRA finetune on C4 repairs the damage, and every experiment fits on a single 40GB A100. On question-answering benchmarks such as MMLU and BoolQ, scores barely change until a large share of layers is gone. Llama-2-70B keeps working until roughly half its layers are removed, then fails abruptly. A cruder rule also works: cut layers from the penultimate one downward, then heal. After healing, it nearly matches the similarity-based method. Next-token loss on C4, in contrast, worsens steadily as pruning grows, and reasoning tasks (GSM8k, HellaSwag) suffer at any pruning level. The authors conclude that current pretraining may underuse deep layers. Shallow layers seem central to storing knowledge, while deep layers matter more for reasoning-style computation.

## Key points
- Method: for each block size n, choose the starting layer ℓ* that minimizes angular distance between layers ℓ and ℓ+n. Remove that block, then optionally heal with QLoRA.
- Llama-2-70B: QA accuracy (MMLU, BoolQ) collapses sharply between roughly 40% and 50% of layers removed. Healed C4 validation loss rises only slowly, even up to about 80% pruning.
- Representations in deeper layers resemble their neighbours more than shallow ones do, except for the final layer. This motivates the simple deep-to-shallow pruning heuristic.
- LoRA rank for healing defaults to 64, with exceptions: Mistral-7B uses rank 4, Llama-2-7B rank 2, and Llama-2-70B rank 8. Smaller ranks raise MMLU but worsen C4 loss, which suggests overfitting at high rank.
- Robustness and baselines:
  - Results for Llama-2-13B are insensitive to the finetuning seed.
  - On Llama-2-7B, similarity-informed pruning beats both random-layer and shallow-layer pruning on MMLU.

## Interesting ideas
- **Similarity as a cheap removal signal.** The angular distance between a block's input and output representations tells you which block can be removed. This needs no gradients or task labels, so it could guide compression of other residual architectures, such as smaller Ukrainian-language models.
- **Prune, then heal.** Removing structure and then running a short parameter-efficient finetune is a low-budget way to shrink a model on one GPU. It also turns pruning into a diagnostic tool for where capabilities live.
- **Evaluate on several kinds of tasks.** Pruning left knowledge-style QA intact but quietly harmed loss and reasoning tasks. Any compression or ablation study should therefore report several task types, not just one headline benchmark.
- **Low LoRA rank as a regularizer.** When healing data is lower quality than the original pretraining data, a smaller adapter rank can protect downstream accuracy. It is worth tracking both task metrics and validation loss to detect this kind of overfitting.
- **Compare against trivial baselines.** Checking the method against random-layer and shallow-layer removal, and against a simple heuristic, shows how much the clever selection rule really adds.

*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

Also on the byline: [Sequoia](https://ror.org/05y5sz883).

📚 Shelves (5): [🖥️ Computer Science](https://md.research.elamur.ai/free/en/shelves/computer-science/) · 9,086 works · [🤖 Artificial Intelligence](https://md.research.elamur.ai/free/en/shelves/computer-science/artificial-intelligence/) · 8,889 works · [🎭 Arts and Humanities](https://md.research.elamur.ai/free/en/shelves/arts-and-humanities/) · 9,296 works · [🗣️ Language and Linguistics](https://md.research.elamur.ai/free/en/shelves/arts-and-humanities/language-and-linguistics/) · 9,206 works · [👁️ Computer Vision and Pattern Recognition](https://md.research.elamur.ai/free/en/shelves/computer-science/computer-vision-and-pattern-recognition/) · 1,282 works

## Cite

Cite as: Andrey Gromov, Kushal Tirumala, Hassan Shapourian, Paolo Glorioso and Daniel A. Roberts. 2024. The Unreasonable Ineffectiveness of the Deeper Layers. arXiv:2403.17887.

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

```bibtex
@misc{gromov2024unreasonable,
  title = {{The Unreasonable Ineffectiveness of the Deeper Layers}},
  author = {Andrey Gromov, Kushal Tirumala, Hassan Shapourian, Paolo Glorioso and Daniel A. Roberts},
  year = {2024},
  doi = {10.48550/arxiv.2403.17887},
  eprint = {2403.17887},
  archivePrefix = {arXiv},
  url = {https://doi.org/10.48550/arxiv.2403.17887}
}
```

- [← previous Seed-TTS: A Family of High-Quality Versatile Speech Generation Models](https://md.research.elamur.ai/free/en/papers/anastassiou-2024-seed-tts-family-high/)
- [next → VeriScore: Evaluating the factuality of verifiable claims in long-form text generation](https://md.research.elamur.ai/free/en/papers/song-2024-veriscore-evaluating-factuality-verifiable-claims/)
- [↑ back to papers of 2024](https://md.research.elamur.ai/free/en/papers/2024/page/4/)

## References

79 references, 30 of them with a card of ours. ↗ leads to the work, «find» searches for one we have no id for yet.

1. 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. nostalgebraist. interpreting gpt: the logit l… [find](https://scholar.google.com/scholar?q=Hugo+Touvron%2C+Louis+Martin%2C+Kevin+Stone%2C+Peter+Albert%2C+Amjad+Almahairi%2C+Yasmine+Babaei%2C+Nikolay+Bashlykov%2C+Soumya+Batra%2C+Prajjwal+Bhargava%2C+Shruti+Bhosale%2C+et+al%2E+Llama+2%3A+Open+foundation+and+fine%2Dtun)
2. Nora Belrose, Zach Furman, Logan Smith, Danny Halawi, Igor Ostrovsky, Lev McKinney, Stella Biderman, and Jacob Steinhardt. [Eliciting latent predictions from transformers with the tuned lens](https://md.research.elamur.ai/free/en/papers/belrose-2023-eliciting-latent-predictions-transformers-tuned/). arXiv preprint arXiv:2303.08112, 2023.
3. Ricky TQ Chen, Yulia Rubanova, Jesse Bettencourt, and David K Duvenaud. Neural ordinary differential equations. Advances in neural information processing systems, 31, 2018. [find](https://scholar.google.com/scholar?q=Ricky+TQ+Chen%2C+Yulia+Rubanova%2C+Jesse+Bettencourt%2C+and+David+K+Duvenaud%2E+Neural+ordinary+differential+equations%2E+Advances+in+neural+information+processing+systems%2C+31%2C+2018%2E)
4. Greg Yang, Dingli Yu, Chen Zhu, and Soufiane Hayou. Tensor programs vi: Feature learning in infinite-depth neural networks. arXiv preprint arXiv:2310.02244, 2023. [find](https://scholar.google.com/scholar?q=Greg+Yang%2C+Dingli+Yu%2C+Chen+Zhu%2C+and+Soufiane+Hayou%2E+Tensor+programs+vi%3A+Feature+learning+in+infinite%2Ddepth+neural+networks%2E+arXiv+preprint+arXiv%3A2310%2E02244%2C+2023%2E)
5. Yann LeCun, John Denker, and Sara Solla. Optimal brain damage. In D. Touretzky, editor, Advances in Neural Information Processing Systems, volume 2. Morgan-Kaufmann, 1989. [find](https://scholar.google.com/scholar?q=Yann+LeCun%2C+John+Denker%2C+and+Sara+Solla%2E+Optimal+brain+damage%2E+In+D%2E+Touretzky%2C+editor%2C+Advances+in+Neural+Information+Processing+Systems%2C+volume+2%2E+Morgan%2DKaufmann%2C+1989%2E)
6. Babak Hassibi and David Stork. Second order derivatives for network pruning: Optimal brain surgeon. In S. Hanson, J. Cowan, and C. Giles, editors, Advances in Neural Information Processing Systems, volume 5. Morgan-Kaufmann, 1992. [find](https://scholar.google.com/scholar?q=Babak+Hassibi+and+David+Stork%2E+Second+order+derivatives+for+network+pruning%3A+Optimal+brain+surgeon%2E+In+S%2E+Hanson%2C+J%2E+Cowan%2C+and+C%2E+Giles%2C+editors%2C+Advances+in+Neural+Information+Processing+Systems%2C+vo)
7. Song Han, Jeff Pool, John Tran, and William Dally. Learning both weights and connections for efficient neural network. Advances in neural information processing systems, 28, 2015. [find](https://scholar.google.com/scholar?q=Song+Han%2C+Jeff+Pool%2C+John+Tran%2C+and+William+Dally%2E+Learning+both+weights+and+connections+for+efficient+neural+network%2E+Advances+in+neural+information+processing+systems%2C+28%2C+2015%2E)
8. Wenlin Chen, James Wilson, Stephen Tyree, Kilian Weinberger, and Yixin Chen. Compressing neural networks with the hashing trick. In International conference on machine learning, pages 2285–2294. PMLR, 2015. [find](https://scholar.google.com/scholar?q=Wenlin+Chen%2C+James+Wilson%2C+Stephen+Tyree%2C+Kilian+Weinberger%2C+and+Yixin+Chen%2E+Compressing+neural+networks+with+the+hashing+trick%2E+In+International+conference+on+machine+learning%2C+pages+2285%E2%80%932294%2E+PMLR%2C)
9. Suraj Srinivas and R Venkatesh Babu. Data-free parameter pruning for deep neural networks. arXiv preprint arXiv:1507.06149, 2015. — [arXiv:1507.06149](https://arxiv.org/abs/1507.06149) [find](https://scholar.google.com/scholar?q=Suraj+Srinivas+and+R+Venkatesh+Babu%2E+Data%2Dfree+parameter+pruning+for+deep+neural+networks%2E+arXiv+preprint+arXiv%3A1507%2E06149%2C+2015%2E+%E2%80%94+%5BarXiv%3A1507%2E06149%5D%28https%3A%2F%2Farxiv%2Eorg%2Fabs%2F1507%2E06149%29)
10. Hao Li, Asim Kadav, Igor Durdanovic, Hanan Samet, and Hans Peter Graf. Pruning filters for efficient convnets. arXiv preprint arXiv:1608.08710, 2016. — [arXiv:1608.08710](https://arxiv.org/abs/1608.08710) [find](https://scholar.google.com/scholar?q=Hao+Li%2C+Asim+Kadav%2C+Igor+Durdanovic%2C+Hanan+Samet%2C+and+Hans+Peter+Graf%2E+Pruning+filters+for+efficient+convnets%2E+arXiv+preprint+arXiv%3A1608%2E08710%2C+2016%2E+%E2%80%94+%5BarXiv%3A1608%2E08710%5D%28https%3A%2F%2Farxiv%2Eorg%2Fabs%2F1608%2E08)
11. Wei Wen, Chunpeng Wu, Yandan Wang, Yiran Chen, and Hai Li. Learning structured sparsity in deep neural networks. Advances in neural information processing systems, 29, 2016. [find](https://scholar.google.com/scholar?q=Wei+Wen%2C+Chunpeng+Wu%2C+Yandan+Wang%2C+Yiran+Chen%2C+and+Hai+Li%2E+Learning+structured+sparsity+in+deep+neural+networks%2E+Advances+in+neural+information+processing+systems%2C+29%2C+2016%2E)
12. Hengyuan Hu, Rui Peng, Yu-Wing Tai, and Chi-Keung Tang. Network trimming: A data-driven neuron pruning approach towards efficient deep architectures. arXiv preprint arXiv:1607.03250, 2016. — [arXiv:1607.03250](https://arxiv.org/abs/1607.03250) [find](https://scholar.google.com/scholar?q=Hengyuan+Hu%2C+Rui+Peng%2C+Yu%2DWing+Tai%2C+and+Chi%2DKeung+Tang%2E+Network+trimming%3A+A+data%2Ddriven+neuron+pruning+approach+towards+efficient+deep+architectures%2E+arXiv+preprint+arXiv%3A1607%2E03250%2C+2016%2E+%E2%80%94+%5BarXiv%3A16)
13. Yihui He, Xiangyu Zhang, and Jian Sun. Channel pruning for accelerating very deep neural networks. In Proceedings of the IEEE international conference on computer vision, pages 1389–1397, 2017. [find](https://scholar.google.com/scholar?q=Yihui+He%2C+Xiangyu+Zhang%2C+and+Jian+Sun%2E+Channel+pruning+for+accelerating+very+deep+neural+networks%2E+In+Proceedings+of+the+IEEE+international+conference+on+computer+vision%2C+pages+1389%E2%80%931397%2C+2017%2E)
14. Gao Huang, Shichen Liu, Laurens Van der Maaten, and Kilian Q Weinberger. Condensenet: An efficient densenet using learned group convolutions. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2752–2761, 2018. [find](https://scholar.google.com/scholar?q=Gao+Huang%2C+Shichen+Liu%2C+Laurens+Van+der+Maaten%2C+and+Kilian+Q+Weinberger%2E+Condensenet%3A+An+efficient+densenet+using+learned+group+convolutions%2E+In+Proceedings+of+the+IEEE+conference+on+computer+vision+a)
15. Kenton Murray and David Chiang. Auto-sizing neural networks: With applications to n-gram language models. arXiv preprint arXiv:1508.05051, 2015. [find](https://scholar.google.com/scholar?q=Kenton+Murray+and+David+Chiang%2E+Auto%2Dsizing+neural+networks%3A+With+applications+to+n%2Dgram+language+models%2E+arXiv+preprint+arXiv%3A1508%2E05051%2C+2015%2E)
16. Abigail See, Minh-Thang Luong, and Christopher D Manning. Compression of neural machine translation models via pruning. arXiv preprint arXiv:1606.09274, 2016. [find](https://scholar.google.com/scholar?q=Abigail+See%2C+Minh%2DThang+Luong%2C+and+Christopher+D+Manning%2E+Compression+of+neural+machine+translation+models+via+pruning%2E+arXiv+preprint+arXiv%3A1606%2E09274%2C+2016%2E)
17. Yoon Kim and Alexander M Rush. arXiv:1606.[07947, 2016](https://md.research.elamur.ai/free/en/papers/kim-2016-sequence-level-knowledge-distillation/).
18. Sequence-level knowledge distillation. Elena Voita, David Talbot, Fedor Moiseev, Rico Sennrich, and Ivan Titov. [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/). arXiv preprint arXiv:1905.09418, 2019.
19. Paul Michel, Omer Levy, and Graham Neubig. [Are sixteen heads really better than one?](https://md.research.elamur.ai/free/en/papers/michel-2019-sixteen-heads-really-better-than/) Advances in neural information processing systems, 32, 2019.
20. Young Jin Kim and Hany Hassan Awadalla. Fastformers: Highly efficient transformer models for natural language understanding. arXiv preprint arXiv:2010.13382, 2020. [find](https://scholar.google.com/scholar?q=Young+Jin+Kim+and+Hany+Hassan+Awadalla%2E+Fastformers%3A+Highly+efficient+transformer+models+for+natural+language+understanding%2E+arXiv+preprint+arXiv%3A2010%2E13382%2C+2020%2E)
21. Angela Fan, Edouard Grave, and Armand Joulin. [Reducing transformer depth on demand with structured dropout](https://md.research.elamur.ai/free/en/papers/fan-2019-reducing-transformer-depth-demand-structured/). arXiv preprint arXiv:1909.11556, 2019.
22. Minjia Zhang and Yuxiong He. Accelerating training of transformer-based language models with progressive layer dropping. Advances in Neural Information Processing Systems, 33:14011–14023, 2020. [find](https://scholar.google.com/scholar?q=Minjia+Zhang+and+Yuxiong+He%2E+Accelerating+training+of+transformer%2Dbased+language+models+with+progressive+layer+dropping%2E+Advances+in+Neural+Information+Processing+Systems%2C+33%3A14011%E2%80%9314023%2C+2020%2E)
23. Chun Fan, Jiwei Li, Xiang Ao, Fei Wu, Yuxian Meng, and Xiaofei Sun. Layer-wise model pruning based on mutual information. arXiv preprint arXiv:2108.12594, 2021. [find](https://scholar.google.com/scholar?q=Chun+Fan%2C+Jiwei+Li%2C+Xiang+Ao%2C+Fei+Wu%2C+Yuxian+Meng%2C+and+Xiaofei+Sun%2E+Layer%2Dwise+model+pruning+based+on+mutual+information%2E+arXiv+preprint+arXiv%3A2108%2E12594%2C+2021%2E)
24. Ananya Harsh Jha, Dirk Groeneveld, Emma Strubell, and Iz Beltagy. Large language model distillation doesn’t need a teacher. arXiv preprint arXiv:2305.14864, 2023. — [arXiv:2305.14864](https://arxiv.org/abs/2305.14864) [find](https://scholar.google.com/scholar?q=Ananya+Harsh+Jha%2C+Dirk+Groeneveld%2C+Emma+Strubell%2C+and+Iz+Beltagy%2E+Large+language+model+distillation+doesn%E2%80%99t+need+a+teacher%2E+arXiv+preprint+arXiv%3A2305%2E14864%2C+2023%2E+%E2%80%94+%5BarXiv%3A2305%2E14864%5D%28https%3A%2F%2Farxiv%2Eor)
25. Hassan Sajjad, Fahim Dalvi, Nadir Durrani, and Preslav Nakov. On the effect of dropping layers of pre-trained transformer models. Computer Speech & Language, 77:101429, 2023. [find](https://scholar.google.com/scholar?q=Hassan+Sajjad%2C+Fahim+Dalvi%2C+Nadir+Durrani%2C+and+Preslav+Nakov%2E+On+the+effect+of+dropping+layers+of+pre%2Dtrained+transformer+models%2E+Computer+Speech+%26+Language%2C+77%3A101429%2C+2023%2E)
26. Wei Liu, Zhiyuan Peng, and Tan Lee. Comflp: Correlation measure based fast search on asr layer pruning. arXiv preprint arXiv:2309.11768, 2023a. — [arXiv:2309.11768](https://arxiv.org/abs/2309.11768) [find](https://scholar.google.com/scholar?q=Wei+Liu%2C+Zhiyuan+Peng%2C+and+Tan+Lee%2E+Comflp%3A+Correlation+measure+based+fast+search+on+asr+layer+pruning%2E+arXiv+preprint+arXiv%3A2309%2E11768%2C+2023a%2E+%E2%80%94+%5BarXiv%3A2309%2E11768%5D%28https%3A%2F%2Farxiv%2Eorg%2Fabs%2F2309%2E11768%29)
27. Lu Hou, Zhiqi Huang, Lifeng Shang, Xin Jiang, Xiao Chen, and Qun Liu. Dynabert: Dynamic bert with adaptive width and depth. Advances in Neural Information Processing Systems, 33: 9782–9793, 2020. [find](https://scholar.google.com/scholar?q=Lu+Hou%2C+Zhiqi+Huang%2C+Lifeng+Shang%2C+Xin+Jiang%2C+Xiao+Chen%2C+and+Qun+Liu%2E+Dynabert%3A+Dynamic+bert+with+adaptive+width+and+depth%2E+Advances+in+Neural+Information+Processing+Systems%2C+33%3A+9782%E2%80%939793%2C+2020%2E)
28. Pratyusha Sharma, Jordan T Ash, and Dipendra Misra. The truth is in there: Improving reasoning in language models with layer-selective rank reduction. arXiv preprint arXiv:2312.13558, 2023. — [arXiv:2312.13558](https://arxiv.org/abs/2312.13558) [find](https://scholar.google.com/scholar?q=Pratyusha+Sharma%2C+Jordan+T+Ash%2C+and+Dipendra+Misra%2E+The+truth+is+in+there%3A+Improving+reasoning+in+language+models+with+layer%2Dselective+rank+reduction%2E+arXiv+preprint+arXiv%3A2312%2E13558%2C+2023%2E+%E2%80%94+%5BarXiv%3A2)
29. Saleh Ashkboos, Maximilian L. Croci, Marcelo Gennari do Nascimento, Torsten Hoefler, and James Hensman. Slicegpt: Compress large language models by deleting rows and columns. arXiv preprint arXiv:2401.15024, 2024. [find](https://scholar.google.com/scholar?q=Saleh+Ashkboos%2C+Maximilian+L%2E+Croci%2C+Marcelo+Gennari+do+Nascimento%2C+Torsten+Hoefler%2C+and+James+Hensman%2E+Slicegpt%3A+Compress+large+language+models+by+deleting+rows+and+columns%2E+arXiv+preprint+arXiv%3A2401)
30. Mengzhou Xia, Zexuan Zhong, and Danqi Chen. [Structured pruning learns compact and accurate models](https://md.research.elamur.ai/free/en/papers/xia-2022-structured-pruning-learns-compact-accurate/). arXiv preprint arXiv:2204.00408, 2022.
31. François Lagunas, Ella Charlaix, Victor Sanh, and Alexander M Rush. [Block pruning for faster transformers](https://md.research.elamur.ai/free/en/papers/lagunas-2021-block-pruning-faster-transformers/). arXiv preprint arXiv:2109.04838, 2021.
32. Xin Men, Mingyu Xu, Qingyu Zhang, Bingning Wang, Hongyu Lin, Yaojie Lu, Xianpei Han, and Weipeng Chen. [Shortgpt: Layers in large language models are more redundant than you expect](https://md.research.elamur.ai/free/en/papers/men-2024-shortgpt-layers-large-language-models/). arXiv preprint arXiv:2403.03853, 2024.
33. Jinze Bai, Shuai Bai, Yunfei Chu, Zeyu Cui, Kai Dang, Xiaodong Deng, Yang Fan, Wenbin Ge, Yu Han, Fei Huang, et al. [Qwen technical report](https://md.research.elamur.ai/free/en/papers/bai-2023-qwen-technical-report/). arXiv preprint arXiv:2309.16609, 2023.
34. Albert Q Jiang, Alexandre Sablayrolles, Arthur Mensch, Chris Bamford, Devendra Singh Chaplot, Diego de las Casas, Florian Bressand, Gianna Lengyel, Guillaume Lample, Lucile Saulnier, et al. Mistral 7b. arXiv preprint arXiv:2310.06825, 2023a. [find](https://scholar.google.com/scholar?q=Albert+Q+Jiang%2C+Alexandre+Sablayrolles%2C+Arthur+Mensch%2C+Chris+Bamford%2C+Devendra+Singh+Chaplot%2C+Diego+de+las+Casas%2C+Florian+Bressand%2C+Gianna+Lengyel%2C+Guillaume+Lample%2C+Lucile+Saulnier%2C+et+al%2E+Mistral+7b)
35. Mojan Javaheripi and Sébastien Bubeck. Phi-2: The surprising power of small language models, Dec 2023. [find](https://scholar.google.com/scholar?q=Mojan+Javaheripi+and+S%C3%A9bastien+Bubeck%2E+Phi%2D2%3A+The+surprising+power+of+small+language+models%2C+Dec+2023%2E)
36. Tim Dettmers, Artidoro Pagnoni, Ari Holtzman, and Luke Zettlemoyer. Qlora: Efficient finetuning of quantized llms. arXiv preprint arXiv:2305.14314, 2023. arXiv preprint Colin Raffel, Noam Shazeer, Adam Roberts, Katherine Lee, Sharan Narang, Michael Matena, Yanqi Zhou, Wei Li, and Peter J Liu. Explor… [find](https://scholar.google.com/scholar?q=Tim+Dettmers%2C+Artidoro+Pagnoni%2C+Ari+Holtzman%2C+and+Luke+Zettlemoyer%2E+Qlora%3A+Efficient+finetuning+of+quantized+llms%2E+arXiv+preprint+arXiv%3A2305%2E14314%2C+2023%2E+arXiv+preprint+Colin+Raffel%2C+Noam+Shazeer%2C+Ada)
37. Dan Hendrycks, Collin Burns, Steven Basart, Andy Zou, Mantas Mazeika, Dawn Song, and Jacob Steinhardt. [Measuring massive multitask language understanding](https://md.research.elamur.ai/free/en/papers/hendrycks-2020-measuring-massive-multitask-language-understanding/). arXiv preprint arXiv:2009.03300, 2020.
38. Christopher Clark, Kenton Lee, Ming-Wei Chang, Tom Kwiatkowski, Michael Collins, and Kristina Toutanova. [Boolq: Exploring the surprising difficulty of natural yes/no questions](https://md.research.elamur.ai/free/en/papers/clark-2019-boolq-exploring-surprising-difficulty-natural/). arXiv preprint arXiv:1905.10044, 2019.
39. Rylan Schaeffer, Brando Miranda, and Sanmi Koyejo. [Are emergent abilities of large language models a mirage?](https://md.research.elamur.ai/free/en/papers/schaeffer-2023-emergent-abilities-large-language-models/) arXiv preprint arXiv:2304.15004, 2023.
40. Karl Cobbe, Vineet Kosaraju, Mohammad Bavarian, Mark Chen, Heewoo Jun, Lukasz Kaiser, Matthias Plappert, Jerry Tworek, Jacob Hilton, Reiichiro Nakano, et al. [Training verifiers to solve math word problems](https://md.research.elamur.ai/free/en/papers/cobbe-2021-training-verifiers-solve-math-word/). arXiv preprint arXiv:2110.14168, 2021.
41. Rowan Zellers, Ari Holtzman, Yonatan Bisk, Ali Farhadi, and Yejin Choi. Hellaswag: Can a machine really finish your sentence? arXiv preprint arXiv:1905.07830, 2019. [find](https://scholar.google.com/scholar?q=Rowan+Zellers%2C+Ari+Holtzman%2C+Yonatan+Bisk%2C+Ali+Farhadi%2C+and+Yejin+Choi%2E+Hellaswag%3A+Can+a+machine+really+finish+your+sentence%3F+arXiv+preprint+arXiv%3A1905%2E07830%2C+2019%2E)
42. Leo Gao, Jonathan Tow, Baber Abbasi, Stella Biderman, Sid Black, Anthony DiPofi, Charles Foster, Laurence Golding, Jeffrey Hsu, Alain Le Noac’h, Haonan Li, Kyle McDonell, Niklas Muennighoff, Chris Ociepa, Jason Phang, Laria Reynolds, Hailey Schoelkopf, Aviya Skowron, Lintang Sutawika, Eric Tang, Ani… [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)
43. Hyung Won Chung, Le Hou, Shayne Longpre, Barret Zoph, Yi Tay, William Fedus, Yunxuan Li, Xuezhi Wang, Mostafa Dehghani, Siddhartha Brahma, et al. [Scaling instruction-finetuned language models](https://md.research.elamur.ai/free/en/papers/chung-2022-scaling-instruction-finetuned-language-models/). Journal of Machine Learning Research, 25(70):1–53, 2024.
44. Mostafa Elhoushi, Akshat Shrivastava, Diana Liskovich, Basil Hosmer, Bram Wasti, Liangzhen Lai, Anas Mahmoud, Bilge Acun, Saurabh Agarwal, Ahmed Roman, et al. Layer skip: Enabling early exit inference and self-speculative decoding. arXiv preprint arXiv:2404.[16710, 2024](https://md.research.elamur.ai/free/en/papers/elhoushi-2024-layerskip-enabling-early-exit-inference/).
45. Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Łukasz Kaiser, and Illia Polosukhin. [Attention is all you need](https://md.research.elamur.ai/free/en/papers/vaswani-2017-attention-all-need/). Advances in neural information processing systems, 30, 2017.
46. Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. [Bert: Pre-training of deep bidirectional transformers for language understanding](https://md.research.elamur.ai/free/en/papers/devlin-2019-bert-pre-training-deep-bidirectional/). arXiv preprint arXiv:1810.04805, 2018.
47. Alec Radford, Jeff Wu, Rewon Child, David Luan, Dario Amodei, and Ilya Sutskever. Language models are unsupervised multitask learners. 2019. URL [https://cdn.openai.com/better-language-models/language_models_are_](https://cdn.openai.com/better-language-models/language_models_are_) unsupervised_multitask_learners.pdf. [↗](https://cdn.openai.com/better-language-models/language_models_are_)
48. Qihuang Zhong, Liang Ding, Juhua Liu, Bo Du, and Dacheng Tao. Can chatgpt understand too? a comparative study on chatgpt and fine-tuned bert. arXiv preprint arXiv:2302.10198, 2023. [find](https://scholar.google.com/scholar?q=Qihuang+Zhong%2C+Liang+Ding%2C+Juhua+Liu%2C+Bo+Du%2C+and+Dacheng+Tao%2E+Can+chatgpt+understand+too%3F+a+comparative+study+on+chatgpt+and+fine%2Dtuned+bert%2E+arXiv+preprint+arXiv%3A2302%2E10198%2C+2023%2E)
49. Kawin Ethayarajh. How contextual are contextualized word representations? comparing the geometry of bert, elmo, and gpt-2 embeddings. arXiv preprint arXiv:1909.00512, 2019. [find](https://scholar.google.com/scholar?q=Kawin+Ethayarajh%2E+How+contextual+are+contextualized+word+representations%3F+comparing+the+geometry+of+bert%2C+elmo%2C+and+gpt%2D2+embeddings%2E+arXiv+preprint+arXiv%3A1909%2E00512%2C+2019%2E)
50. Alexei Baevski, Yuhao Zhou, Abdelrahman Mohamed, and Michael Auli. [wav2vec 2.0: A framework for self-supervised learning of speech representations](https://md.research.elamur.ai/free/en/papers/baevski-2020-wav2vec-2-0-framework-self/). Advances in neural information processing systems, 33:12449–12460, 2020.
51. Geoffrey Hinton, Oriol Vinyals, and Jeff Dean. [Distilling the knowledge in a neural network](https://md.research.elamur.ai/free/en/papers/hinton-2015-distilling-knowledge-neural-network/). arXiv preprint arXiv:1503.02531, 2015.
52. Yuxian Gu, Li Dong, Furu Wei, and Minlie Huang. Knowledge distillation of large language models. arXiv preprint arXiv:2306.[08543, 2023](https://md.research.elamur.ai/free/en/papers/gu-2023-minillm-policy-distillation-large-language/).
53. Xiaoqi Jiao, Yichun Yin, Lifeng Shang, Xin Jiang, Xiao Chen, Linlin Li, Fang Wang, and Qun Liu. Tinybert: Distilling bert for natural language understanding. arXiv preprint arXiv:1909.10351, 2019. [find](https://scholar.google.com/scholar?q=Xiaoqi+Jiao%2C+Yichun+Yin%2C+Lifeng+Shang%2C+Xin+Jiang%2C+Xiao+Chen%2C+Linlin+Li%2C+Fang+Wang%2C+and+Qun+Liu%2E+Tinybert%3A+Distilling+bert+for+natural+language+understanding%2E+arXiv+preprint+arXiv%3A1909%2E10351%2C+2019%2E)
54. Shuohang Wang, Yang Liu, Yichong Xu, Chenguang Zhu, and Michael Zeng. [Want to reduce labeling cost? gpt-3 can help](https://md.research.elamur.ai/free/en/papers/wang-2021-want-reduce-labeling-cost-gpt/). arXiv preprint arXiv:2108.13487, 2021.
55. Ronen Eldan and Yuanzhi Li. [Tinystories: How small can language models be and still speak coherent english?](https://md.research.elamur.ai/free/en/papers/eldan-2023-tinystories-small-language-models-still/) arXiv preprint arXiv:2305.07759, 2023.
56. Yuanzhi Li, Sébastien Bubeck, Ronen Eldan, Allie Del Giorno, Suriya Gunasekar, and Yin Tat Lee. [Textbooks are all you need ii: phi-1.5 technical report](https://md.research.elamur.ai/free/en/papers/li-2023-textbooks-all-need-ii-phi/). arXiv preprint arXiv:2309.05463, 2023a.
57. Suriya Gunasekar, Yi Zhang, Jyoti Aneja, Caio César Teodoro Mendes, Allie Del Giorno, Sivakanth Gopi, Mojan Javaheripi, Piero Kauffmann, Gustavo de Rosa, Olli Saarikivi, et al. Textbooks are all you need. arXiv preprint arXiv:2306.11644, 2023. [find](https://scholar.google.com/scholar?q=Suriya+Gunasekar%2C+Yi+Zhang%2C+Jyoti+Aneja%2C+Caio+C%C3%A9sar+Teodoro+Mendes%2C+Allie+Del+Giorno%2C+Sivakanth+Gopi%2C+Mojan+Javaheripi%2C+Piero+Kauffmann%2C+Gustavo+de+Rosa%2C+Olli+Saarikivi%2C+et+al%2E+Textbooks+are+all+you+n)
58. Yao Fu, Hao Peng, Litu Ou, Ashish Sabharwal, and Tushar Khot. [Specializing smaller language models towards multi-step reasoning](https://md.research.elamur.ai/free/en/papers/fu-2023-specializing-smaller-language-models-towards/). arXiv preprint arXiv:2301.12726, 2023.
59. Cheng-Yu Hsieh, Chun-Liang Li, Chih-Kuan Yeh, Hootan Nakhost, Yasuhisa Fujii, Alexander Ratner, Ranjay Krishna, Chen-Yu Lee, and Tomas Pfister. [Distilling step-by-step! outperforming larger language models with less training data and smaller model sizes](https://md.research.elamur.ai/free/en/papers/hsieh-2023-distilling-step-step-outperforming-larger/). arXiv preprint arXiv:2305.02301, 2023.
60. Yuxin Jiang, Chunkit Chan, Mingyang Chen, and Wei Wang. Lion: Adversarial distillation of closed-source large language model. arXiv preprint arXiv:2305.12870, 2023b. — [arXiv:2305.12870](https://arxiv.org/abs/2305.12870) [find](https://scholar.google.com/scholar?q=Yuxin+Jiang%2C+Chunkit+Chan%2C+Mingyang+Chen%2C+and+Wei+Wang%2E+Lion%3A+Adversarial+distillation+of+closed%2Dsource+large+language+model%2E+arXiv+preprint+arXiv%3A2305%2E12870%2C+2023b%2E+%E2%80%94+%5BarXiv%3A2305%2E12870%5D%28https%3A%2F%2Farxiv)
61. Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. [Lora: Low-rank adaptation of large language models](https://md.research.elamur.ai/free/en/papers/hu-2021-lora-low-rank-adaptation-large/). arXiv preprint arXiv:2106.09685, 2021.
62. Yixiao Li, Yifan Yu, Chen Liang, Pengcheng He, Nikos Karampatziakis, Weizhu Chen, and Tuo Zhao. Loftq: Lora-fine-tuning-aware quantization for large language models. arXiv preprint arXiv:2310.08659, 2023b. — [arXiv:2310.08659](https://arxiv.org/abs/2310.08659) [find](https://scholar.google.com/scholar?q=Yixiao+Li%2C+Yifan+Yu%2C+Chen+Liang%2C+Pengcheng+He%2C+Nikos+Karampatziakis%2C+Weizhu+Chen%2C+and+Tuo+Zhao%2E+Loftq%3A+Lora%2Dfine%2Dtuning%2Daware+quantization+for+large+language+models%2E+arXiv+preprint+arXiv%3A2310%2E08659%2C+2)
63. Qingru Zhang, Minshuo Chen, Alexander Bukharin, Pengcheng He, Yu Cheng, Weizhu Chen, and Tuo Zhao. Adaptive budget allocation for parameter-efficient fine-tuning. arXiv preprint arXiv:2303.[10512, 2023](https://md.research.elamur.ai/free/en/papers/zhang-2023-adalora-adaptive-budget-allocation-parameter/).
64. Yaniv Leviathan, Matan Kalman, and Yossi Matias. Fast inference from transformers via speculative decoding. In International Conference on Machine Learning, pages 19274–19286. PMLR, 2023. [find](https://scholar.google.com/scholar?q=Yaniv+Leviathan%2C+Matan+Kalman%2C+and+Yossi+Matias%2E+Fast+inference+from+transformers+via+speculative+decoding%2E+In+International+Conference+on+Machine+Learning%2C+pages+19274%E2%80%9319286%2E+PMLR%2C+2023%2E)
65. Tianle Cai, Yuhong Li, Zhengyang Geng, Hongwu Peng, Jason D Lee, Deming Chen, and Tri Dao. [Medusa: Simple llm inference acceleration framework with multiple decoding heads](https://md.research.elamur.ai/free/en/papers/cai-2024-medusa-simple-llm-inference-acceleration/). arXiv preprint arXiv:2401.10774, 2024.
66. Kevin Meng, David Bau, Alex Andonian, and Yonatan Belinkov. Locating and editing factual associations in gpt. Advances in Neural Information Processing Systems, 35:17359–17372, 2022. [find](https://scholar.google.com/scholar?q=Kevin+Meng%2C+David+Bau%2C+Alex+Andonian%2C+and+Yonatan+Belinkov%2E+Locating+and+editing+factual+associations+in+gpt%2E+Advances+in+Neural+Information+Processing+Systems%2C+35%3A17359%E2%80%9317372%2C+2022%2E)
67. Damai Dai, Li Dong, Yaru Hao, Zhifang Sui, Baobao Chang, and Furu Wei. Knowledge neurons in pretrained transformers. arXiv preprint arXiv:2104.08696, 2021. [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+Knowledge+neurons+in+pretrained+transformers%2E+arXiv+preprint+arXiv%3A2104%2E08696%2C+2021%2E)
68. Peter Hase, Mohit Bansal, Been Kim, and Asma Ghandeharioun. Does localization inform editing? surprising differences in causality-based localization vs. knowledge editing in language models. arXiv preprint arXiv:2301.04213, 2023. [find](https://scholar.google.com/scholar?q=Peter+Hase%2C+Mohit+Bansal%2C+Been+Kim%2C+and+Asma+Ghandeharioun%2E+Does+localization+inform+editing%3F+surprising+differences+in+causality%2Dbased+localization+vs%2E+knowledge+editing+in+language+models%2E+arXiv+pre)
69. Mor Geva, Jasmijn Bastings, Katja Filippova, and Amir Globerson. Dissecting recall of factual associations in auto-regressive language models. arXiv preprint arXiv:2304.14767, 2023. [find](https://scholar.google.com/scholar?q=Mor+Geva%2C+Jasmijn+Bastings%2C+Katja+Filippova%2C+and+Amir+Globerson%2E+Dissecting+recall+of+factual+associations+in+auto%2Dregressive+language+models%2E+arXiv+preprint+arXiv%3A2304%2E14767%2C+2023%2E)
70. Alexander Yom Din, Taelin Karidi, Leshem Choshen, and Mor Geva. Jump to conclusions: Shortcutting transformers with linear transformations. arXiv preprint arXiv:2303.09435, 2023. — [arXiv:2303.09435](https://arxiv.org/abs/2303.09435) [find](https://scholar.google.com/scholar?q=Alexander+Yom+Din%2C+Taelin+Karidi%2C+Leshem+Choshen%2C+and+Mor+Geva%2E+Jump+to+conclusions%3A+Shortcutting+transformers+with+linear+transformations%2E+arXiv+preprint+arXiv%3A2303%2E09435%2C+2023%2E+%E2%80%94+%5BarXiv%3A2303%2E09435%5D%28)
71. Wes Gurnee and Max Tegmark. Language models represent space and time. arXiv preprint arXiv:2310.02207, 2023. [find](https://scholar.google.com/scholar?q=Wes+Gurnee+and+Max+Tegmark%2E+Language+models+represent+space+and+time%2E+arXiv+preprint+arXiv%3A2310%2E02207%2C+2023%2E)
72. Elena Voita, Javier Ferrando, and Christoforos Nalmpantis. Neurons in large language models: Dead, n-gram, positional. arXiv preprint arXiv:2309.04827, 2023. [find](https://scholar.google.com/scholar?q=Elena+Voita%2C+Javier+Ferrando%2C+and+Christoforos+Nalmpantis%2E+Neurons+in+large+language+models%3A+Dead%2C+n%2Dgram%2C+positional%2E+arXiv+preprint+arXiv%3A2309%2E04827%2C+2023%2E)
73. Zichang Liu, Jue Wang, Tri Dao, Tianyi Zhou, Binhang Yuan, Zhao Song, Anshumali Shrivastava, Ce Zhang, Yuandong Tian, Christopher Re, et al. Deja vu: Contextual sparsity for efficient llms at inference time. In International Conference on Machine Learning, pages 22137–22176. PMLR, 2023b. [find](https://scholar.google.com/scholar?q=Zichang+Liu%2C+Jue+Wang%2C+Tri+Dao%2C+Tianyi+Zhou%2C+Binhang+Yuan%2C+Zhao+Song%2C+Anshumali+Shrivastava%2C+Ce+Zhang%2C+Yuandong+Tian%2C+Christopher+Re%2C+et+al%2E+Deja+vu%3A+Contextual+sparsity+for+efficient+llms+at+inferenc)
74. Abhishek Panigrahi, Nikunj Saunshi, Haoyu Zhao, and Sanjeev Arora. Task-specific skill localization in fine-tuned language models. arXiv preprint arXiv:2302.06600, 2023. — [arXiv:2302.06600](https://arxiv.org/abs/2302.06600) [find](https://scholar.google.com/scholar?q=Abhishek+Panigrahi%2C+Nikunj+Saunshi%2C+Haoyu+Zhao%2C+and+Sanjeev+Arora%2E+Task%2Dspecific+skill+localization+in+fine%2Dtuned+language+models%2E+arXiv+preprint+arXiv%3A2302%2E06600%2C+2023%2E+%E2%80%94+%5BarXiv%3A2302%2E06600%5D%28https%3A%2F%2Fa)
75. [Thomas Wolf, Lysandre Debut, Victor Sanh, Julien Chaumond, Clement Delangue, Anthony Moi, Pierric Cistac, Tim Rault, Rémi Louf, Morgan Funtowicz, Joe Davison, Sam Shleifer, Patrick von Platen, Clara Ma, Yacine Jernite, Julien Plu, Canwen Xu, Teven Le Scao, Sylvain Gugger, Mariama Drame, Quentin Lhoe…](https://md.research.elamur.ai/free/en/papers/wolf-2020-transformers-state-art-natural-language/)
76. Colin Raffel, Noam Shazeer, Adam Roberts, Katherine Lee, Sharan Narang, Michael Matena, Yanqi Zhou, Wei Li, and Peter J. Liu. [Exploring the limits of transfer learning with a unified text-to-text transformer](https://md.research.elamur.ai/free/en/papers/raffel-2019-exploring-limits-transfer-learning-unified/). arXiv e-prints, 2019.
77. Sourab Mangrulkar, Sylvain Gugger, Lysandre Debut, Younes Belkada, Sayak Paul, and Benjamin Bossan. Peft: State-of-the-art parameter-efficient fine-tuning methods. [https://github.com/huggingface/peft](https://github.com/huggingface/peft), 2022. [↗](https://github.com/huggingface/peft)
78. Ariel N Lee, Cole J Hunter, and Nataniel Ruiz. Platypus: Quick, cheap, and powerful refinement of llms. arXiv preprint arXiv:2308.07317, 2023. [find](https://scholar.google.com/scholar?q=Ariel+N+Lee%2C+Cole+J+Hunter%2C+and+Nataniel+Ruiz%2E+Platypus%3A+Quick%2C+cheap%2C+and+powerful+refinement+of+llms%2E+arXiv+preprint+arXiv%3A2308%2E07317%2C+2023%2E)
79. Tim Dettmers, Mike Lewis, Younes Belkada, and Luke Zettlemoyer. Llm. int8 (): 8-bit matrix multiplication for transformers at scale. arXiv preprint arXiv:2208.07339, 2022. [find](https://scholar.google.com/scholar?q=Tim+Dettmers%2C+Mike+Lewis%2C+Younes+Belkada%2C+and+Luke+Zettlemoyer%2E+Llm%2E+int8+%28%29%3A+8%2Dbit+matrix+multiplication+for+transformers+at+scale%2E+arXiv+preprint+arXiv%3A2208%2E07339%2C+2022%2E)

*Updated 24 October 2026*