---
title: "Chain-of-Thought Compression Should Not Be Blind: V-Skip for"
description: "The authors show that this causes \"Visual Amnesia\": the compressor discards words… 48.2% accuracy. 35 references, cited by 3 papers here."
authors: ["Dongxu Zhang", "Yiding Sun", "Cheng Tan", "Wenbiao Yan", "Ning Yang", "Jihua Zhu", "Haijun Zhang"]
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.887/"
doi: "10.18653/v1/2026.acl-long.887"
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/zhang-2026-chain-thought-compression-should-not/"
---
[Elamur Research](https://md.research.elamur.ai/free/en/) · [this page in HTML](https://research.elamur.ai/free/en/papers/zhang-2026-chain-thought-compression-should-not/)

[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/5/) / Chain-of-Thought Compression Should Not Be Blind: V-Skip for Efficient Multimodal Reasoning via Dual-Path Anchoring

# Chain-of-Thought Compression Should Not Be Blind: V-Skip for Efficient Multimodal Reasoning via Dual-Path Anchoring

**Authors:** [Dongxu Zhang](https://md.research.elamur.ai/free/en/authors/dongxu-zhang/), [Yiding Sun](https://md.research.elamur.ai/free/en/authors/yiding-sun/), Cheng Tan, Wenbiao Yan, Ning Yang, Jihua Zhu and Haijun Zhang · **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.887/) · [PDF](https://aclanthology.org/2026.acl-long.887.pdf) · [DOI 10.18653/v1/2026.acl-long.887](https://doi.org/10.18653/v1/2026.acl-long.887) · [OpenAlex](https://openalex.org/W7166889240)

**Citations:** 0 citations (OpenAlex, 2026-10-14) · [3](https://md.research.elamur.ai/free/en/papers/zhang-2026-chain-thought-compression-should-not/cited-by/) in our library, of the 149,439 papers whose bibliographies we read

## Summary
Long Chain-of-Thought (CoT) rationales make Multimodal Large Language Models (MLLMs) more accurate, but they also make them slow and memory-hungry. Text-oriented compressors such as TokenSkip, LLMLingua-2 and AS-CoT remove tokens that look predictable from the surrounding text. The authors show that this causes "Visual Amnesia": the compressor discards words that are easy to guess linguistically but carry image evidence, such as colours or numbers read from a document. The model then hallucinates.

V-Skip treats pruning as a Visual-Anchored Information Bottleneck (VA-IB) problem. A dual-path gate scores each token twice, once by linguistic surprisal and once by cross-modal attention flow. A token survives if either path marks it as important.

Training runs offline in three steps:
1. A VLM generates full CoTs.
2. CoTs that lead to wrong answers are filtered out, and the rest are pruned to a target ratio γ.
3. A LoRA adapter is trained on the compressed chains, so inference needs no extra scoring.

On Qwen2-VL and Llama-3.2 models, V-Skip runs about 2.9× faster with little accuracy loss. It keeps visual details intact and reduces hallucination, which makes it relevant to cheaper deployment of grounded multimodal reasoning.

## Key points
- **Speed and accuracy:** V-Skip gives a reported 2.9× speedup with negligible accuracy loss. The gains over baselines are stated inconsistently: the abstract claims more than 30% on DocVQA, while the conclusion claims more than 45% on fine-grained visual tasks.
- **Visual Attribute Retention Rate (VARR), MMMU, γ = 0.5:** V-Skip keeps 89.4% of colour, 86.3% of shape and 91.2% of object tokens. LLMLingua-2 keeps 42.5%, 55.1% and 64.8%; ASCoT keeps 48.2%, 58.7% and 69.5%.
- **Hallucination on POPE:**
  - Yes-Ratio: V-Skip 51.2%, uncompressed baseline 50.4%, ASCoT 66.8%, LLMLingua-2 64.5%.
  - F1: V-Skip 88.9, full CoT 89.1, ASCoT 77.1, LLMLingua-2 78.9.
- **Ablation on MMMU:** combining the two scores as a union gives 48.2% accuracy with a 1.3% syntax error rate.
  - Text-only scoring: 32.4% accuracy.
  - Vision-only scoring: 40.5% accuracy, but an 11.2% syntax error rate.
  - Intersection of both scores: 30.1% accuracy.
- **Limitations:** the method relies on the base model's cross-modal alignment and requires offline distillation, which adds training cost.

*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

- [Harbin Institute of Technology](https://md.research.elamur.ai/free/en/universities/harbin-institute-of-technology/) — University, Harbin, China · 158 works · 28,841 citations
- [Chinese Academy of Sciences](https://md.research.elamur.ai/free/en/government/chinese-academy-of-sciences/) — Government body, Beijing, China · 194 works · 23,906 citations
- [Shanghai Artificial Intelligence Laboratory](https://md.research.elamur.ai/free/en/labs/shanghai-artificial-intelligence-laboratory/) — Research institute, Shanghai, China · 117 works · 11,281 citations
- [Institute of Automation](https://md.research.elamur.ai/free/en/labs/institute-of-automation/) — Research institute, Beijing, China · 70 works · 14,251 citations
- [Xi'an Jiaotong University](https://md.research.elamur.ai/free/en/universities/xi-an-jiaotong-university/) — University, Xi'an, China · 22 works · 36,046 citations
- [Xidian University](https://md.research.elamur.ai/free/en/universities/xidian-university/) — University, Xi'an, China · 14 works · 1,357 citations
- [University of Science and Technology Beijing](https://md.research.elamur.ai/free/en/universities/university-of-science-and-technology-beijing/) — University, Beijing, China · 8 works · 315 citations

📚 Shelves (5): [🖥️ Computer Science](https://md.research.elamur.ai/free/en/shelves/computer-science/) · 16,667 works · [🤖 Artificial Intelligence](https://md.research.elamur.ai/free/en/shelves/computer-science/artificial-intelligence/) · 16,194 works · [🎭 Arts and Humanities](https://md.research.elamur.ai/free/en/shelves/arts-and-humanities/) · 16,039 works · [🗣️ Language and Linguistics](https://md.research.elamur.ai/free/en/shelves/arts-and-humanities/language-and-linguistics/) · 15,891 works · [👁️ Computer Vision and Pattern Recognition](https://md.research.elamur.ai/free/en/shelves/computer-science/computer-vision-and-pattern-recognition/) · 2,869 works

🧭 Topics (2): [🧮 Reasoning](https://md.research.elamur.ai/free/en/topics/reasoning/) · 1,426 works · [🖼️ Vision & language](https://md.research.elamur.ai/free/en/topics/multimodal/) · 657 works

## Cite

Cite as: Dongxu Zhang, Yiding Sun, Cheng Tan, Wenbiao Yan, Ning Yang, Jihua Zhu and Haijun Zhang. 2026. Chain-of-Thought Compression Should Not Be Blind: V-Skip for Efficient Multimodal Reasoning via Dual-Path Anchoring. 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{zhang2026chain,
  title = {{Chain-of-Thought Compression Should Not Be Blind: V-Skip for Efficient Multimodal Reasoning via Dual-Path Anchoring}},
  author = {Dongxu Zhang, Yiding Sun, Cheng Tan, Wenbiao Yan, Ning Yang, Jihua Zhu and Haijun Zhang},
  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.887},
  url = {https://doi.org/10.18653/v1/2026.acl-long.887}
}
```

- [← previous Can ChatGPT Really Understand Modern Chinese Poetry?](https://md.research.elamur.ai/free/en/papers/wang-2026-can-chatgpt-really-understand-modern/)
- [next → Chronological Thinking in Full-Duplex Spoken Dialogue Language Models](https://md.research.elamur.ai/free/en/papers/wu-2026-chronological-thinking-full-duplex-spoken/)
- [↑ back to papers of 2026](https://md.research.elamur.ai/free/en/papers/2026/page/5/)

## References

35 references, 14 of them with a card of ours. ↗ leads to the work, «find» searches for one we have no id for yet.

1. Alexander A Alemi, Ian Fischer, Joshua V Dillon, and Kevin Murphy. 2016. Deep variational information bottleneck. arXiv preprint arXiv:1612.00410. — [arXiv:1612.00410](https://arxiv.org/abs/1612.00410) [find](https://scholar.google.com/scholar?q=Alexander+A+Alemi%2C+Ian+Fischer%2C+Joshua+V+Dillon%2C+and+Kevin+Murphy%2E+2016%2E+Deep+variational+information+bottleneck%2E+arXiv+preprint+arXiv%3A1612%2E00410%2E+%E2%80%94+%5BarXiv%3A1612%2E00410%5D%28https%3A%2F%2Farxiv%2Eorg%2Fabs%2F1612%2E00410)
2. Maciej Besta, Nils Blach, Ales Kubicek, Robert Gerstenberger, Michal Podstawski, Lukas Gianinazzi, Joanna Gajda, Tomasz Lehmann, Hubert Niewiadomski, Piotr Nyczyk, and 1 others. 2024. Graph of thoughts: Solving elaborate problems with large language models. In Proceedings of the AAAI conference on a… [find](https://scholar.google.com/scholar?q=Maciej+Besta%2C+Nils+Blach%2C+Ales+Kubicek%2C+Robert+Gerstenberger%2C+Michal+Podstawski%2C+Lukas+Gianinazzi%2C+Joanna+Gajda%2C+Tomasz+Lehmann%2C+Hubert+Niewiadomski%2C+Piotr+Nyczyk%2C+and+1+others%2E+2024%2E+Graph+of+thought)
3. Yu-Neng Chuang, Tianwei Xing, Chia-Yuan Chang, Zirui Liu, Xun Chen, and Xia Hu. 2024. Learning to compress prompt in natural language formats. arXiv preprint arXiv:2402.18700. [find](https://scholar.google.com/scholar?q=Yu%2DNeng+Chuang%2C+Tianwei+Xing%2C+Chia%2DYuan+Chang%2C+Zirui+Liu%2C+Xun+Chen%2C+and+Xia+Hu%2E+2024%2E+Learning+to+compress+prompt+in+natural+language+formats%2E+arXiv+preprint+arXiv%3A2402%2E18700%2E)
4. Tri Dao, Dan Fu, Stefano Ermon, Atri Rudra, and Christopher Ré. 2022. Flashattention: Fast and memory-efficient exact attention with io-awareness. Advances in neural information processing systems, 35:16344–16359. [find](https://scholar.google.com/scholar?q=Tri+Dao%2C+Dan+Fu%2C+Stefano+Ermon%2C+Atri+Rudra%2C+and+Christopher+R%C3%A9%2E+2022%2E+Flashattention%3A+Fast+and+memory%2Defficient+exact+attention+with+io%2Dawareness%2E+Advances+in+neural+information+processing+systems%2C+35)
5. Tim Dettmers, Mike Lewis, Younes Belkada, and Luke Zettlemoyer. 2022. Gpt3. int8 (): 8-bit matrix multiplication for transformers at scale. Advances in neural information processing systems, 35:30318– 30332. [find](https://scholar.google.com/scholar?q=Tim+Dettmers%2C+Mike+Lewis%2C+Younes+Belkada%2C+and+Luke+Zettlemoyer%2E+2022%2E+Gpt3%2E+int8+%28%29%3A+8%2Dbit+matrix+multiplication+for+transformers+at+scale%2E+Advances+in+neural+information+processing+systems%2C+35%3A30318%E2%80%93)
6. Abhimanyu Dubey, Abhinav Jauhri, Abhinav Pandey, Abhishek Kadian, Ahmad Al-Dahle, Aiesha Letman, Akhil Mathur, Alan Schelten, Amy Yang, Angela Fan, 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 e-prints, pages arXiv–2407.
7. Huiqiang Jiang, Qianhui Wu, Chin-Yew Lin, Yuqing Yang, and Lili Qiu. 2023. [Llmlingua: Compressing prompts for accelerated inference of large language models](https://md.research.elamur.ai/free/en/papers/jiang-2023-llmlingua-compressing-prompts-accelerated-inference/). arXiv preprint arXiv:2310.05736.
8. Woosuk Kwon, Zhuohan Li, Siyuan Zhuang, Ying Sheng, Lianmin Zheng, Cody Hao Yu, Joseph Gonzalez, Hao Zhang, and Ion Stoica. 2023. [Efficient memory management for large language model serving with pagedattention](https://md.research.elamur.ai/free/en/papers/kwon-2023-efficient-memory-management-large-language/). In Proceedings of the 29th symposium on operating systems principles, pages 611–626.
9. Feng Li, Renrui Zhang, Hao Zhang, Yuanhan Zhang, Bo Li, Wei Li, Zejun Ma, and Chunyuan Li. 2024. Llava-next-interleave: Tackling multi-image, video, and 3d in large multimodal models. arXiv preprint arXiv:2407.07895. [find](https://scholar.google.com/scholar?q=Feng+Li%2C+Renrui+Zhang%2C+Hao+Zhang%2C+Yuanhan+Zhang%2C+Bo+Li%2C+Wei+Li%2C+Zejun+Ma%2C+and+Chunyuan+Li%2E+2024%2E+Llava%2Dnext%2Dinterleave%3A+Tackling+multi%2Dimage%2C+video%2C+and+3d+in+large+multimodal+models%2E+arXiv+preprint+a)
10. Yifan Li, Yifan Du, Kun Zhou, Jinpeng Wang, Wayne Xin Zhao, and Ji-Rong Wen. 2023. [Evaluating object hallucination in large vision-language models](https://md.research.elamur.ai/free/en/papers/li-2023-evaluating-object-hallucination-large-vision/). arXiv preprint arXiv:2305.10355.
11. Liang Lin, Zhihao Xu, Junhao Dong, Jian Zhao, Yuchen Yuan, Guibin Zhang, Miao Yu, Yiming Zhang, Zhengtao Yao, Huahui Yi, and 1 others. 2025. Or-thalign: Orthogonal subspace decomposition for non-interfering multi-objective alignment. arXiv preprint arXiv:2509.24610. — [arXiv:2509.24610](https://arxiv.org/abs/2509.24610) [find](https://scholar.google.com/scholar?q=Liang+Lin%2C+Zhihao+Xu%2C+Junhao+Dong%2C+Jian+Zhao%2C+Yuchen+Yuan%2C+Guibin+Zhang%2C+Miao+Yu%2C+Yiming+Zhang%2C+Zhengtao+Yao%2C+Huahui+Yi%2C+and+1+others%2E+2025%2E+Or%2Dthalign%3A+Orthogonal+subspace+decomposition+for+non%2Dinter)
12. Liang Lin, Miao Yu, Kaiwen Luo, Yibo Zhang, Lilan Peng, Dexian Wang, Xuehai Tang, Yuanhe Zhang, Xikang Yang, Zhenhong Zhou, and 1 others. 2026. Hidden in the noise: Unveiling backdoors in audio llms alignment through latent acoustic pattern triggers. In Proceedings of the AAAI Conference on Artifici… [find](https://scholar.google.com/scholar?q=Liang+Lin%2C+Miao+Yu%2C+Kaiwen+Luo%2C+Yibo+Zhang%2C+Lilan+Peng%2C+Dexian+Wang%2C+Xuehai+Tang%2C+Yuanhe+Zhang%2C+Xikang+Yang%2C+Zhenhong+Zhou%2C+and+1+others%2E+2026%2E+Hidden+in+the+noise%3A+Unveiling+backdoors+in+audio+llms+a)
13. Minesh Mathew, Dimosthenis Karatzas, and CV Jawahar. 2021. [Docvqa: A dataset for vqa on document images](https://md.research.elamur.ai/free/en/papers/mathew-2020-docvqa-dataset-vqa-document-images/). In Proceedings of the IEEE/CVF winter conference on applications of computer vision, pages 2200–2209.
14. Catherine Olsson, Nelson Elhage, Neel Nanda, Nicholas Joseph, Nova DasSarma, Tom Henighan, Ben Mann, Amanda Askell, Yuntao Bai, Anna Chen, and 1 others. 2022. [In-context learning and induction heads](https://md.research.elamur.ai/free/en/papers/olsson-2022-context-learning-induction-heads/). arXiv preprint arXiv:2209.11895.
15. Aaron van den Oord, Yazhe Li, and Oriol Vinyals. 2018. [Representation learning with contrastive predictive coding](https://md.research.elamur.ai/free/en/papers/oord-2018-representation-learning-contrastive-predictive-coding/). arXiv preprint arXiv:1807.03748.
16. Kawin Ethayarajh. 2019. How contextual are contextualized word representations. Comparing the geometry of BERT, ELMo, and GPT-2 Embeddings, 2. [find](https://scholar.google.com/scholar?q=Kawin+Ethayarajh%2E+2019%2E+How+contextual+are+contextualized+word+representations%2E+Comparing+the+geometry+of+BERT%2C+ELMo%2C+and+GPT%2D2+Embeddings%2C+2%2E)
17. Zhuoshi Pan, Qianhui Wu, Huiqiang Jiang, Menglin Xia, Xufang Luo, Jue Zhang, Qingwei Lin, Victor Rühle, Yuqing Yang, Chin-Yew Lin, and 1 others. 2024. [Llmlingua-2: Data distillation for efficient and faithful task-agnostic prompt compression](https://md.research.elamur.ai/free/en/papers/pan-2024-llmlingua-2-data-distillation-efficient/). arXiv preprint arXiv:2403.12968.
18. Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, Weizhu Chen, and 1 others. 2022. [Lora: Low-rank adaptation of large language models](https://md.research.elamur.ai/free/en/papers/hu-2021-lora-low-rank-adaptation-large/). ICLR, 1(2):3.
19. Anna Rohrbach, Lisa Anne Hendricks, Kaylee Burns, Trevor Darrell, and Kate Saenko. 2018. [Object hallucination in image captioning](https://md.research.elamur.ai/free/en/papers/rohrbach-2018-object-hallucination-image-captioning/). arXiv preprint arXiv:1809.02156.
20. Noam Shazeer. 2019. Fast transformer decoding: One write-head is all you need. arXiv preprint arXiv:1911.02150. [find](https://scholar.google.com/scholar?q=Noam+Shazeer%2E+2019%2E+Fast+transformer+decoding%3A+One+write%2Dhead+is+all+you+need%2E+arXiv+preprint+arXiv%3A1911%2E02150%2E)
21. Naftali Tishby, Fernando C Pereira, and William Bialek. 2000. The information bottleneck method. arXiv preprint physics/0004057. [find](https://scholar.google.com/scholar?q=Naftali+Tishby%2C+Fernando+C+Pereira%2C+and+William+Bialek%2E+2000%2E+The+information+bottleneck+method%2E+arXiv+preprint+physics%2F0004057%2E)
22. Lei Wang, Yi Hu, Jiabang He, Xing Xu, Ning Liu, Hui Liu, and Heng Tao Shen. 2024a. T-sciq: Teaching multimodal chain-of-thought reasoning via large language model signals for science question answering. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 38, pages 19162–19170. [find](https://scholar.google.com/scholar?q=Lei+Wang%2C+Yi+Hu%2C+Jiabang+He%2C+Xing+Xu%2C+Ning+Liu%2C+Hui+Liu%2C+and+Heng+Tao+Shen%2E+2024a%2E+T%2Dsciq%3A+Teaching+multimodal+chain%2Dof%2Dthought+reasoning+via+large+language+model+signals+for+science+question+answerin)
23. Peng Wang, Shuai Bai, Sinan Tan, Shijie Wang, Zhihao Fan, Jinze Bai, Keqin Chen, Xuejing Liu, Jialin Wang, Wenbin Ge, and 1 others. 2024b. [Qwen2-vl: Enhancing vision-language model’s perception of the world at any resolution](https://md.research.elamur.ai/free/en/papers/wang-2024-qwen2-vl-enhancing-vision/). arXiv preprint arXiv:2409.12191.
24. Yaoting Wang, Shengqiong Wu, Yuecheng Zhang, Shuicheng Yan, Ziwei Liu, Jiebo Luo, and Hao Fei. 2025. Multimodal chain-of-thought reasoning: A comprehensive survey. arXiv preprint arXiv:2503.12605. — [arXiv:2503.12605](https://arxiv.org/abs/2503.12605) [find](https://scholar.google.com/scholar?q=Yaoting+Wang%2C+Shengqiong+Wu%2C+Yuecheng+Zhang%2C+Shuicheng+Yan%2C+Ziwei+Liu%2C+Jiebo+Luo%2C+and+Hao+Fei%2E+2025%2E+Multimodal+chain%2Dof%2Dthought+reasoning%3A+A+comprehensive+survey%2E+arXiv+preprint+arXiv%3A2503%2E12605%2E+%E2%80%94+%5B)
25. Dongxu Zhang, Hongqiang Lin, Yiding Sun, Pengyu Wang, Qirui Wang, Ning Yang, and Jihua Zhu. 2026a. Not all queries need deep thought: Coficot for adaptive coarse-to-fine stateful refinement. arXiv preprint arXiv:2603.08251. — [arXiv:2603.08251](https://arxiv.org/abs/2603.08251) [find](https://scholar.google.com/scholar?q=Dongxu+Zhang%2C+Hongqiang+Lin%2C+Yiding+Sun%2C+Pengyu+Wang%2C+Qirui+Wang%2C+Ning+Yang%2C+and+Jihua+Zhu%2E+2026a%2E+Not+all+queries+need+deep+thought%3A+Coficot+for+adaptive+coarse%2Dto%2Dfine+stateful+refinement%2E+arXiv+pre)
26. Dongxu Zhang, Yiding Sun, Pengcheng Li, Yumou Liu, Hongqiang Lin, Haoran Xu, Xiaoxuan Mu, Liang Lin, Wenbiao Yan, Ning Yang, and 1 others. 2026b. Pointcot: A multi-modal benchmark for explicit 3d geometric reasoning. arXiv preprint arXiv:2602.23945. — [arXiv:2602.23945](https://arxiv.org/abs/2602.23945) [find](https://scholar.google.com/scholar?q=Dongxu+Zhang%2C+Yiding+Sun%2C+Pengcheng+Li%2C+Yumou+Liu%2C+Hongqiang+Lin%2C+Haoran+Xu%2C+Xiaoxuan+Mu%2C+Liang+Lin%2C+Wenbiao+Yan%2C+Ning+Yang%2C+and+1+others%2E+2026b%2E+Pointcot%3A+A+multi%2Dmodal+benchmark+for+explicit+3d+geom)
27. Dongxu Zhang, Ning Yang, Jihua Zhu, Jinnan Yang, Miao Xin, and Baoliang Tian. 2025. Ascot: An adaptive self-correction chain-of-thought method for late-stage fragility in llms. arXiv preprint arXiv:2508.05282. — [arXiv:2508.05282](https://arxiv.org/abs/2508.05282) [find](https://scholar.google.com/scholar?q=Dongxu+Zhang%2C+Ning+Yang%2C+Jihua+Zhu%2C+Jinnan+Yang%2C+Miao+Xin%2C+and+Baoliang+Tian%2E+2025%2E+Ascot%3A+An+adaptive+self%2Dcorrection+chain%2Dof%2Dthought+method+for+late%2Dstage+fragility+in+llms%2E+arXiv+preprint+arXiv%3A25)
28. Zhuosheng Zhang, Aston Zhang, Mu Li, Hai Zhao, George Karypis, and Alex Smola. 2023. [Multimodal chain-of-thought reasoning in language models](https://md.research.elamur.ai/free/en/papers/zhang-2023-multimodal-chain-thought-reasoning-language/). arXiv preprint arXiv:2302.00923.
29. Peijun Zhu, Ning Yang, Jiayu Wei, Jinghang Wu, and Haijun Zhang. 2025. Breaking the moe llm trilemma: Dynamic expert clustering with structured compression. arXiv preprint arXiv:2510.02345. — [arXiv:2510.02345](https://arxiv.org/abs/2510.02345) [find](https://scholar.google.com/scholar?q=Peijun+Zhu%2C+Ning+Yang%2C+Jiayu+Wei%2C+Jinghang+Wu%2C+and+Haijun+Zhang%2E+2025%2E+Breaking+the+moe+llm+trilemma%3A+Dynamic+expert+clustering+with+structured+compression%2E+arXiv+preprint+arXiv%3A2510%2E02345%2E+%E2%80%94+%5BarXiv%3A2)
30. Heming Xia, Chak Tou Leong, Wenjie Wang, Yongqi Li, and Wenjie Li. 2025. [Tokenskip: Controllable chain-of-thought compression in llms](https://md.research.elamur.ai/free/en/papers/xia-2025-tokenskip-controllable-chain-thought-compression/). arXiv preprint arXiv:2502.12067.
31. Ning Yang, Hai Lin, Yibo Liu, Baoliang Tian, Guoqing Liu, and Haijun Zhang. 2025. Token-importance guided direct preference optimization. arXiv preprint arXiv:2505.19653. — [arXiv:2505.19653](https://arxiv.org/abs/2505.19653) [find](https://scholar.google.com/scholar?q=Ning+Yang%2C+Hai+Lin%2C+Yibo+Liu%2C+Baoliang+Tian%2C+Guoqing+Liu%2C+and+Haijun+Zhang%2E+2025%2E+Token%2Dimportance+guided+direct+preference+optimization%2E+arXiv+preprint+arXiv%3A2505%2E19653%2E+%E2%80%94+%5BarXiv%3A2505%2E19653%5D%28https%3A%2F%2F)
32. Zhengyuan Yang, Zhe Gan, Jianfeng Wang, Xiaowei Hu, Yumao Lu, Zicheng Liu, and Lijuan Wang. 2022. An empirical study of gpt-3 for few-shot knowledge-based vqa. In Proceedings of the AAAI conference on artificial intelligence, volume 36, pages 3081–3089. [find](https://scholar.google.com/scholar?q=Zhengyuan+Yang%2C+Zhe+Gan%2C+Jianfeng+Wang%2C+Xiaowei+Hu%2C+Yumao+Lu%2C+Zicheng+Liu%2C+and+Lijuan+Wang%2E+2022%2E+An+empirical+study+of+gpt%2D3+for+few%2Dshot+knowledge%2Dbased+vqa%2E+In+Proceedings+of+the+AAAI+conference+on)
33. Junyi Yao, Yijiang Liu, Zhen Dong, Mingfei Guo, Helan Hu, Kurt Keutzer, Li Du, Daquan Zhou, and Shanghang Zhang. 2024. Promptcot: Align prompt distribution via adapted chain-of-thought. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 7027–7037. [find](https://scholar.google.com/scholar?q=Junyi+Yao%2C+Yijiang+Liu%2C+Zhen+Dong%2C+Mingfei+Guo%2C+Helan+Hu%2C+Kurt+Keutzer%2C+Li+Du%2C+Daquan+Zhou%2C+and+Shanghang+Zhang%2E+2024%2E+Promptcot%3A+Align+prompt+distribution+via+adapted+chain%2Dof%2Dthought%2E+In+Proceedings)
34. Shunyu Yao, Dian Yu, Jeffrey Zhao, Izhak Shafran, Thomas L Griffiths, Yuan Cao, and Karthik Narasimhan. 2023. Tree of thoughts: Deliberate problem solving with large language models, 2023. URL https://arxiv.org/abs/2305.10601, 3:1. [↗](https://arxiv.org/abs/2305.10601)
35. Xiang Yue, Yuansheng Ni, Kai Zhang, Tianyu Zheng, Ruoqi Liu, Ge Zhang, Samuel Stevens, Dongfu Jiang, Weiming Ren, Yuxuan Sun, and 1 others. 2024. [Mmmu: A massive multi-discipline multimodal understanding and reasoning benchmark for expert agi](https://md.research.elamur.ai/free/en/papers/yue-2023-mmmu-massive-multi-discipline-multimodal/). In Proceedings of the IEEE/CVF Conference on Computer Vi…

*Updated 1 November 2026*