Elamur Research⌕
☰
home / papers / 2026 / From Signal Degradation to Computation Collapse: Uncovering the Two Failure Modes of LLM Quantization

From Signal Degradation to Computation Collapse: Uncovering the Two Failure Modes of LLM Quantization

Paper · 2026 · Chenxi Zhou, Pengfei Cao, Jiang Li et al.

Summary

The paper asks why weight-only post-training quantization (PTQ) of LLMs behaves acceptably at 4-bit but breaks down almost completely at 2-bit. It tests whether the 2-bit "performance cliff" is simply a stronger version of 4-bit damage or a different kind of failure. The authors run a mechanistic study of factual recall on Pararel. Llama-3.1-8B is the main model, with checks on Qwen3-8B, Mistral-7B-Instruct-v0.3 and Gemma-2-9B-it. Models are quantized mainly with GPTQ, and AWQ serves as a cross-check. The analysis uses logit lens probing, causal activation patching, zeroing ablation and layer-by-layer quantization tests.

They find two qualitatively different modes. In Signal Degradation (4-bit), the internal circuitry still works, but accumulated noise weakens the correct answer. In Computation Collapse (2-bit), key components stop functioning and the signal is destroyed in early layers. Training-free fixes largely repair the first mode but not the second, which implies 2-bit models need structural reconstruction such as fine-tuning. This gives practitioners a diagnostic lens for choosing between compensation and retraining.

Key points

  • Under 4-bit, the correct answer usually slips only slightly in rank, often staying in the top-5. Under 2-bit, its rank falls into the thousands and outputs degenerate into frequent tokens such as "the" or ".".
  • A two-stage repair raises 4-bit accuracy on the Failure Subset from 0% to 75.19% (Llama3.1-8B), 81.26% (Mistral-7B), 79.88% (Qwen3-8B) and 64.08% (Gemma2-9B). The two stages are:
    • Source protection: first two layers kept in 8-bit (about 4.25 average bits), or kurtosis-based protection (about 4.1 bits).
    • Peak-layer logit amplification: the layer with the highest confidence has its output logits scaled up.
  • On Llama3.1-8B, quantizing only layers 0–1 to 2-bit lowers Robust Subset accuracy to 41.65%. The remaining FP16 layers cannot recover the lost signal.
  • When 2-bit layers receive high-precision inputs, layer-output cosine similarity stays above 0.9 until the first 2-bit layer and then collapses. Neither protection strategies nor EORA low-rank compensation rescue 2-bit models.
  • The study is limited to weight-only quantization and factual recall. MMLU and GSM8K checks appear in an appendix.

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

🏛️Chinese Academy of SciencesBeijing · China194 works · 23,906 citations🔬Institute of AutomationBeijing · China70 works · 14,251 citations🎓University of Chinese Academy of SciencesBeijing · China147 works · 8,480 citations🎓Inner Mongolia UniversityHohhot · China3 works · 1 citation

 Fields, topics and concepts

📚 Shelves 5The field of science this work belongs to (OpenAlex fields).Computer Science · 16,667 worksArtificial Intelligence · 16,194 worksArts and Humanities · 16,039 worksLanguage and Linguistics · 15,891 worksComputer Vision and Pattern Recognition · 2,869 works

🧭 Topics 2Research subjects and languages the work is about.Language models & pretraining · 3,320 worksEfficient models & compression · 242 works

Cite

Cite as Chenxi Zhou, Pengfei Cao, Jiang Li, Bohan Yu, Jinyu Ye, Jun Zhao and Kang Liu. 2026. From Signal Degradation to Computation Collapse: Uncovering the Two Failure Modes of LLM Quantization. Findings of the Association for Computational Linguistics: ACL 2026.

The paper itself, not this page. BibTeX from the library's record of the work. DOI 10.18653/v1/2026.findings-acl.1162

@inproceedings{zhou2026signal,
  title = {{From Signal Degradation to Computation Collapse: Uncovering the Two Failure Modes of LLM Quantization}},
  author = {Chenxi Zhou, Pengfei Cao, Jiang Li, Bohan Yu, Jinyu Ye, Jun Zhao and Kang Liu},
  booktitle = {Findings of the Association for Computational Linguistics: ACL 2026},
  year = {2026},
  doi = {10.18653/v1/2026.findings-acl.1162},
  url = {https://doi.org/10.18653/v1/2026.findings-acl.1162}
}

References

32 references, 8 of them with a card of ours. ↗ leads to the work, «find» searches for one we have no id for yet.

  1. Ting-Yun Chang, Muru Zhang, Jesse Thomason, and Robin Jia. 2025. Why Do Some Inputs Break Low-Bit LLM Quantization? arXiv. find
  2. Vishnu Kabir Chhabra and Mohammad Mahdi Khalili. 2025. Towards Understanding and Improving Refusal in Compressed Models via Mechanistic Interpretability. arXiv preprint. find
  3. Karl Cobbe, Vineet Kosaraju, Mohammad Bavarian, Mark Chen, Heewoo Jun, Lukasz Kaiser, Matthias Plappert, Jerry Tworek, Jacob Hilton, Reiichiro Nakano, Christopher Hesse, and John Schulman. 2021. Training Verifiers to Solve Math Word Problems. arXiv preprint. ArXiv:2110.14168 [cs].
  4. Tim Dettmers, Ruslan Svirschevski, Vage Egiazarian, Denis Kuznedelev, Elias Frantar, Saleh Ashkboos, Alexander Borzunov, Torsten Hoefler, and Dan Alistarh. 2023. SpQR: A Sparse-Quantized Representation for Near-Lossless LLM Weight Compression. arXiv. find
  5. Zhen Dong, Zhewei Yao, Daiyaan Arfeen, Amir Gholami, Michael W Mahoney, and Kurt Keutzer. 2020. HAWQ-V2: Hessian Aware trace-Weighted Quantization of Neural Networks. In Advances in Neural Information Processing Systems, volume 33, pages 18518–18529. Curran Associates, Inc. find
  6. Razvan-Gabriel Dumitru, Vikas Yadav, Rishabh Maheshwary, Paul Ioan Clotan, Sathwik Tejaswi Madhusudhan, and Mihai Surdeanu. 2025. Variable Layerwise Quantization: A Simple and Effective Approach to Quantize LLMs. In Findings of the Association for Computational Linguistics: ACL 2025, pages 534–550,… find
  7. Yanai Elazar, Nora Kassner, Shauli Ravfogel, Abhilasha Ravichander, Eduard Hovy, Hinrich Schütze, and Yoav Goldberg. 2021. Measuring and Improving Consistency in Pretrained Language Models. Transactions of the Association for Computational Linguistics, 9:1012–1031. find
  8. Elias Frantar, Saleh Ashkboos, Torsten Hoefler, and Dan Alistarh. 2023. GPTQ: Accurate Post-Training Quantization for Generative Pre-trained Transformers. In The International Conference on Learning Representations. find
  9. Yao Fu, Xianxuan Long, Runchao Li, Haotian Yu, Mu Sheng, Xiaotian Han, Yu Yin, and Pan Li. 2025. Quantized but Deceptive? A Multi-Dimensional Truthfulness Evaluation of Quantized LLMs. arXiv. ArXiv:2508.19432 [cs]. find
  10. Mor Geva, Roei Schuster, Jonathan Berant, and Omer Levy. 2021. Transformer Feed-Forward Layers Are Key-Value Memories. arXiv. find
  11. Aaron Grattafiori, Abhimanyu Dubey, Abhinav Jauhri, Abhinav Pandey, Abhishek Kadian, Ahmad Al-Dahle, Aiesha Letman, Akhil Mathur, Alan Schelten, Alex Vaughan, Amy Yang, Angela Fan, Anirudh Goyal, Anthony Hartshorn, Aobo Yang, Archi Mitra, Archie Sravankumar, Artem Korenev, Arthur Hinsvark, and 181 o… — The Llama 3 Herd of Models
  12. Stefan Heimersheim and Neel Nanda. 2024. How to use and interpret activation patching. arXiv preprint. find
  13. Dan Hendrycks, Collin Burns, Steven Basart, Andy Zou, Mantas Mazeika, Dawn Song, and Jacob Steinhardt. 2021. Measuring Massive Multitask Language Understanding. arXiv preprint. ArXiv:2009.03300 [cs].
  14. Albert Q. Jiang, Alexandre Sablayrolles, Arthur Mensch, Chris Bamford, Devendra Singh Chaplot, Diego de las Casas, Florian Bressand, Gianna Lengyel, Guillaume Lample, Lucile Saulnier, Lélio Renard Lavaud, Marie-Anne Lachaux, Pierre Stock, Teven Le Scao, Thibaut Lavril, Thomas Wang, Timothée Lacroix,… find
  15. Renren Jin, Jiangcun Du, Wuwei Huang, Wei Liu, Jian Luan, Bin Wang, and Deyi Xiong. 2024. A Comprehensive Evaluation of Quantization Strategies for Large Language Models. arXiv. find
  16. Simon Kornblith, Mohammad Norouzi, Honglak Lee, and Geoffrey Hinton. 2019. Similarity of Neural Network Representations Revisited. In Proceedings of the 36th International Conference on Machine Learning, pages 3519–3529. PMLR. ISSN: 2640-3498.
  17. Shiyao Li, Xuefei Ning, Luning Wang, Tengxuan Liu, Xiangsheng Shi, Shengen Yan, Guohao Dai, Huazhong Yang, and Yu Wang. 2024. Evaluating Quantized Large Language Models. arXiv. find
  18. Ji Lin, Jiaming Tang, Haotian Tang, Shang Yang, Wei-Ming Chen, Wei-Chen Wang, Guangxuan Xiao, Xingyu Dang, Chuang Gan, and Song Han. 2024. AWQ: Activation-aware Weight Quantization for On-Device LLM Compression and Acceleration. Proceedings of Machine Learning and Systems, 6:87– 100. find
  19. Ruikang Liu, Yuxuan Sun, Manyi Zhang, Haoli Bai, Xianzhi Yu, Tiezheng Yu, Chun Yuan, and Lu Hou. 2025a. Quantization Hurts Reasoning? An Empirical Study on Quantized Reasoning Models. arXiv. find
  20. Shih-Yang Liu, Maksim Khadkevich, Nai Chit Fung, Charbel Sakr, Chao-Han Huck Yang, Chien-Yi Wang, Saurav Muralidharan, Hongxu Yin, Kwang-Ting Cheng, Jan Kautz, Yu-Chiang Frank Wang, Pavlo Molchanov, and Min-Hung Chen. 2024. EoRA: Fine-tuning-free Compensation for Compressed LLM with Eigenspace Low-R… find
  21. Zechun Liu, Changsheng Zhao, Igor Fedorov, Bilge Soran, Dhruv Choudhary, Raghuraman Krishnamoorthi, Vikas Chandra, Yuandong Tian, and Tijmen Blankevoort. 2025b. SpinQuant: LLM quantization with learned rotations. arXiv. ArXiv:2405.16406 [cs]. — [arXiv:2405.16406](https://arxiv.org/abs/2405.16406) find
  22. Accurate and Efficient Post-Training Quantization for Large Language Models. In Proceedings of the 40th International Conference on Machine Learning, pages 38087–38099. PMLR. find
  23. He Xiao, Qingyao Yang, Dirui Xie, Wendong Xu, Wenyong Zhou, Haobo Liu, Zhengwu Liu, and Ngai Wong. 2025. Exploring Layer-wise Information Effectiveness for Post-Training Quantization in Small Language Models. arXiv preprint.
  24. An Yang, Anfeng Li, Baosong Yang, Beichen Zhang, Binyuan Hui, Bo Zheng, Bowen Yu, Chang Gao, Chengen Huang, Chenxu Lv, Chujie Zheng, Dayiheng Liu, Fan Zhou, Fei Huang, Feng Hu, Hao Ge, Haoran Wei, Huan Lin, Jialong Tang, and 41 others. 2025. Qwen3 Technical Report. arXiv preprint. ArXiv:2505.09388 […
  25. Feng Zhang, Yanbin Liu, Weihua Li, Jie Lv, Xiaodan Wang, and Quan Bai. 2025a. Towards Superior Quantization Accuracy: A Layer-sensitive Approach. arXiv preprint. find
  26. Zhiwei Zhang, Fali Wang, Xiaomin Li, Zongyu Wu, Xianfeng Tang, Hui Liu, Qi He, Wenpeng Yin, and Suhang Wang. 2025b. Catastrophic Failure of LLM Unlearning via Quantization. arXiv. find
  27. Kevin Meng, David Bau, Alex Andonian, and Yonatan Belinkov. 2022. Locating and Editing Factual Associations in GPT. In Advances in Neural Information Processing Systems, volume 35, pages 17359–17372. find
  28. Satya Sai Srinath Namburi, Makesh Sreedhar, Srinath Srinivasan, and Frederic Sala. 2023. Investigating the Impact of Compression on Parametric Knowledge in Language Models. nostalgebraist. 2020. interpreting GPT: the logit lens. find
  29. Colin Raffel, Noam Shazeer, Adam Roberts, Katherine Lee, Sharan Narang, Michael Matena, Yanqi Zhou, Wei Li, and Peter J. Liu. 2020. Exploring the Limits of Transfer Learning with a Unified Text-to-Text Transformer. Journal of Machine Learning Research, 21(140):1–67.
  30. Gemma Team, Morgane Riviere, Shreya Pathak, Pier Giuseppe Sessa, Cassidy Hardin, Surya Bhupatiraju, Léonard Hussenot, Thomas Mesnard, Bobak Shahriari, Alexandre Ramé, Johan Ferret, Peter Liu, Pouya Tafti, Abe Friesen, Michelle Casbon, Sabela Ramos, Ravin Kumar, Charline Le Lan, Sammy Jerome, and 179… find
  31. Albert Tseng, Jerry Chee, Qingyao Sun, Volodymyr Kuleshov, and Christopher De Sa. 2024. QuIP\#: Even Better LLM Quantization with Hadamard Incoherence and Lattice Codebooks. arXiv preprint. ArXiv:2402.04396 [cs].
  32. Guangxuan Xiao, Ji Lin, Mickael Seznec, Hao Wu, Julien Demouth, and Song Han. 2023. SmoothQuant: find

Updated

↑ back to top

This page in Markdown → https://md.research.elamur.ai/free/en/papers/zhou-2026-signal-degradation-computation-collapse-uncovering/