---
title: "From Redundancy to Relevance — paper by Zhang et al. (2025)"
description: "The authors study this \"information flow\", meaning how strongly system, image and user-prompt… 27 references, cited by 7 papers here."
authors: ["Xiaofeng Zhang", "Yihao Quan", "Chen Shen", "Xiaosong Yuan", "Shaotian Yan", "Liang Xie", "Wenxiao Wang", "Chaochen Gu", "Hao Tang", "Jieping Ye"]
year: 2025
venue: "Proceedings of the 2025 Conference of the Nations of the Americas Chapter of the Association for Computational Linguistics: Human Language Technologies (Volume 1: Long Papers)"
source: "https://aclanthology.org/2025.naacl-long.115/"
doi: "10.18653/v1/2025.naacl-long.115"
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-2025-redundancy-relevance-information-flow-lvlms/"
---
[Elamur Research](https://md.research.elamur.ai/free/en/) · [this page in HTML](https://research.elamur.ai/free/en/papers/zhang-2025-redundancy-relevance-information-flow-lvlms/)

[home](https://md.research.elamur.ai/free/en/) / [papers](https://md.research.elamur.ai/free/en/papers/) / [2025](https://md.research.elamur.ai/free/en/papers/2025/page/8/) / From Redundancy to Relevance: Information Flow in LVLMs Across Reasoning Tasks

# From Redundancy to Relevance: Information Flow in LVLMs Across Reasoning Tasks

**Authors:** [Xiaofeng Zhang](https://md.research.elamur.ai/free/en/authors/xiaofeng-zhang/), [Yihao Quan](https://md.research.elamur.ai/free/en/authors/yihao-quan/), [Chen Shen](https://md.research.elamur.ai/free/en/authors/chen-shen/), [Xiaosong Yuan](https://md.research.elamur.ai/free/en/authors/xiaosong-yuan/), [Shaotian Yan](https://md.research.elamur.ai/free/en/authors/shaotian-yan/), Liang Xie, [Wenxiao Wang](https://md.research.elamur.ai/free/en/authors/wenxiao-wang/), [Chaochen Gu](https://md.research.elamur.ai/free/en/authors/chaochen-gu/), [Hao Tang](https://md.research.elamur.ai/free/en/authors/hao-tang/) and [Jieping Ye](https://md.research.elamur.ai/free/en/authors/jieping-ye/) · **Year:** 2025 · **Venue:** [Proceedings of the 2025 Conference of the Nations of the Americas Chapter of the Association for Computational Linguistics: Human Language Technologies (Volume 1: Long Papers)](https://md.research.elamur.ai/free/en/venues/naacl/) ([NAACL 2025](https://md.research.elamur.ai/free/en/venues/naacl/2025/)) · **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/2025.naacl-long.115/) · [PDF](https://aclanthology.org/2025.naacl-long.115.pdf) · [DOI 10.18653/v1/2025.naacl-long.115](https://doi.org/10.18653/v1/2025.naacl-long.115) · [OpenAlex](https://openalex.org/W4411119489)

**Citations:** 5 citations (OpenAlex, 2026-10-14) · [7](https://md.research.elamur.ai/free/en/papers/zhang-2025-redundancy-relevance-information-flow-lvlms/cited-by/) in our library, of the 149,439 papers whose bibliographies we read

## Summary
Large vision-language models (LVLMs) turn each image into hundreds of tokens before passing it to an LLM. How those visual tokens actually shape the generated answer has been unclear. The authors study this "information flow", meaning how strongly system, image and user-prompt tokens influence answer tokens, layer by layer. They use two complementary tools. Forward-pass attention scores show where the model attends. LLaVA-CAM, a gradient-based saliency method adapted from Smooth-CAM, traces backward which image regions contribute to the answer. Image information concentrates on relevant regions in shallow layers and then scatters in deeper ones. The authors call the point where visual tokens stop mattering the "information flow cliff layer". This cliff arrives earlier for complex reasoning, as in ScienceQA, and later for simpler tasks like image captioning. To test the claim, they remove image tokens from the cliff layer onward in six LVLMs. Accuracy holds steady or even improves on POPE, ScienceQA, TextVQA and CHAIR. This indicates that late-layer visual tokens are largely redundant, which matters for interpretability and for efficient inference.

## Key points
- After layer 3 in LLaVA, image tokens receive under 2% of the answer's attention compared with system and prompt tokens. Attention to image and user tokens rises sharply again at the final, 32nd layer.
- Six models were tested: LLaVA1.5, LLaVA1.6, Qwen-VL, Intern-VL, InstructBLIP and Shikra. Their projectors differ (Linear, MLP, cross-attention), yet all show the same pattern. The authors conclude the behavior comes from the underlying LLM rather than the projector.
- Truncation results for LLaVA1.5: POPE accuracy rose from 84.70 to 85.51, ScienceQA from 65.00 to 66.48, and TextVQA from 59.34 to 60.02.
- Truncation results for Intern-VL: POPE accuracy rose from 94.01 to 94.58 and TextVQA from 57.00 to 58.18.
- For captioning, information flow in LLaVA1.5 diverges only around layer 30. Code is available at github.com/zhangbaijin/From-Redundancy-to-Relevance.

*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

- [Peking University](https://md.research.elamur.ai/free/en/universities/peking-university/) — University, Beijing, China · 315 works · 43,267 citations
- [Alibaba Group](https://md.research.elamur.ai/free/en/companies/alibaba-group/) — Company, Hangzhou, China · 201 works · 17,173 citations
- [Shanghai Jiao Tong University](https://md.research.elamur.ai/free/en/universities/shanghai-jiao-tong-university/) — University, Shanghai, China · 212 works · 19,544 citations
- [Beijing Jiaotong University](https://md.research.elamur.ai/free/en/universities/beijing-jiaotong-university/) — University, Beijing, China · 28 works · 1,754 citations

📚 Shelves (5): [🖥️ 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 · [👁️ Computer Vision and Pattern Recognition](https://md.research.elamur.ai/free/en/shelves/computer-science/computer-vision-and-pattern-recognition/) · 3,486 works

🧭 Topics (1): [🧮 Reasoning](https://md.research.elamur.ai/free/en/topics/reasoning/) · 1,685 works

## Cite

Cite as: Xiaofeng Zhang, Yihao Quan, Chen Shen, Xiaosong Yuan, Shaotian Yan, Liang Xie, Wenxiao Wang, Chaochen Gu, Hao Tang and Jieping Ye. 2025. From Redundancy to Relevance: Information Flow in LVLMs Across Reasoning Tasks. Proceedings of the 2025 Conference of the Nations of the Americas Chapter of the Association for Computational Linguistics: Human Language Technologies (Volume 1: Long Papers).

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

```bibtex
@inproceedings{zhang2025redundancy,
  title = {{From Redundancy to Relevance: Information Flow in LVLMs Across Reasoning Tasks}},
  author = {Xiaofeng Zhang, Yihao Quan, Chen Shen, Xiaosong Yuan, Shaotian Yan, Liang Xie, Wenxiao Wang, Chaochen Gu, Hao Tang and Jieping Ye},
  booktitle = {Proceedings of the 2025 Conference of the Nations of the Americas Chapter of the Association for Computational Linguistics: Human Language Technologies (Volume 1: Long Papers)},
  year = {2025},
  doi = {10.18653/v1/2025.naacl-long.115},
  url = {https://doi.org/10.18653/v1/2025.naacl-long.115}
}
```

- [← previous from Benign import Toxic: Jailbreaking the Language Model via Adversarial Metaphors](https://md.research.elamur.ai/free/en/papers/yan-2025-benign-import-toxic-jailbreaking-language/)
- [next → GenAI Content Detection Task 2: AI vs. Human – Academic Essay Authenticity Challenge](https://md.research.elamur.ai/free/en/papers/chowdhury-2025-genai-content-detection-task-2/)
- [↑ back to papers of 2025](https://md.research.elamur.ai/free/en/papers/2025/page/8/)

## References

27 references, 15 of them with a card of ours. ↗ leads to the work, «find» searches for one we have no id for yet.

1. Jinze Bai, Shuai Bai, Shusheng Yang, Shijie Wang, Sinan Tan, Peng Wang, Junyang Lin, Chang Zhou, and Jingren Zhou. 2023. [Qwen-vl: A frontier large vision-language model with versatile abilities](https://md.research.elamur.ai/free/en/papers/bai-2023-qwen-vl-versatile-vision-language/). arXiv preprint arXiv:2308.12966.
2. Jun Chen, Deyao Zhu, Xiaoqian Shen, Xiang Li, Zechun Liu, Pengchuan Zhang, Raghuraman Krishnamoorthi, Vikas Chandra, Yunyang Xiong, and Mohamed Elhoseiny. 2023a. [Minigpt-v2: large language model as a unified interface for vision-language multi-task learning](https://md.research.elamur.ai/free/en/papers/chen-2023-minigpt-v2-large-language-model/). arXiv preprint arXiv:2310.09478.
3. Keqin Chen, Zhao Zhang, Weili Zeng, Richong Zhang, Feng Zhu, and Rui Zhao. 2023b. [Shikra: Unleashing multimodal llm’s referential dialogue magic](https://md.research.elamur.ai/free/en/papers/chen-2023-shikra-unleashing-multimodal-llm-s/). arXiv preprint arXiv:2306.15195.
4. Liang Chen, Haozhe Zhao, Tianyu Liu, Shuai Bai, Junyang Lin, Chang Zhou, and Baobao Chang. 2024a. [An image is worth 1/2 tokens after layer 2: Plug-and-play inference acceleration for large vision-language models](https://md.research.elamur.ai/free/en/papers/chen-2024-image-worth-1-2-tokens/). 18th European Conference on Computer Vision ECCV 2024.
5. Zhe Chen, Jiannan Wu, Wenhai Wang, Weijie Su, Guo Chen, Sen Xing, Muyan Zhong, Qinglong Zhang, Xizhou Zhu, Lewei Lu, et al. 2024b. [Internvl: Scaling up vision foundation models and aligning for generic visual-linguistic tasks](https://md.research.elamur.ai/free/en/papers/chen-2023-internvl-scaling-up-vision/). In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern…
6. Xiangxiang Chu, Limeng Qiao, Xinyang Lin, Shuang Xu, Yang Yang, Yiming Hu, Fei Wei, Xinyu Zhang, Bo Zhang, Xiaolin Wei, et al. 2023. Mobilevlm: A fast, reproducible and strong vision language assistant for mobile devices. arXiv preprint arXiv:2312.16886. [find](https://scholar.google.com/scholar?q=Xiangxiang+Chu%2C+Limeng+Qiao%2C+Xinyang+Lin%2C+Shuang+Xu%2C+Yang+Yang%2C+Yiming+Hu%2C+Fei+Wei%2C+Xinyu+Zhang%2C+Bo+Zhang%2C+Xiaolin+Wei%2C+et+al%2E+2023%2E+Mobilevlm%3A+A+fast%2C+reproducible+and+strong+vision+language+assistan)
7. Kevin Clark, Urvashi Khandelwal, Omer Levy, and Christopher D Manning. 2019. What does bert look at? an analysis of bert’s attention. arXiv preprint arXiv:1906.04341. [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+arXiv+preprint+arXiv%3A1906%2E04341%2E)
8. Damai Dai, Li Dong, Yaru Hao, Zhifang Sui, Baobao Chang, and Furu Wei. 2021. Knowledge neurons in pretrained transformers. arXiv preprint arXiv:2104.08696. [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+2021%2E+Knowledge+neurons+in+pretrained+transformers%2E+arXiv+preprint+arXiv%3A2104%2E08696%2E)
9. Wenliang Dai, Junnan Li, Dongxu Li, Anthony Meng Huat Tiong, Junqi Zhao, Weisheng Wang, Boyang Li, Pascale N Fung, and Steven Hoi. 2024. [Instructblip: Towards general-purpose vision-language models with instruction tuning](https://md.research.elamur.ai/free/en/papers/dai-2023-instructblip-towards-general-purpose-vision/). Advances in Neural Information Processing Systems, 36.
10. Jiaming Han, Renrui Zhang, Wenqi Shao, Peng Gao, Peng Xu, Han Xiao, Kaipeng Zhang, Chris Liu, Song Wen, Ziyu Guo, et al. 2023. Imagebind-llm: Multi-modality instruction tuning. arXiv preprint arXiv:2309.03905. — [arXiv:2309.03905](https://arxiv.org/abs/2309.03905) [find](https://scholar.google.com/scholar?q=Jiaming+Han%2C+Renrui+Zhang%2C+Wenqi+Shao%2C+Peng+Gao%2C+Peng+Xu%2C+Han+Xiao%2C+Kaipeng+Zhang%2C+Chris+Liu%2C+Song+Wen%2C+Ziyu+Guo%2C+et+al%2E+2023%2E+Imagebind%2Dllm%3A+Multi%2Dmodality+instruction+tuning%2E+arXiv+preprint+arXiv%3A23)
11. John Hewitt and Christopher D Manning. 2019. [A structural probe for finding syntax in word representations](https://md.research.elamur.ai/free/en/papers/hewitt-2019-structural-probe-finding-syntax-word/). In Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 (Long and Short Papers), pages 4129–…
12. Qidong Huang, Xiaoyi Dong, Pan Zhang, Bin Wang, Conghui He, Jiaqi Wang, Dahua Lin, Weiming Zhang, and Nenghai Yu. 2024. [Opera: Alleviating hallucination in multi-modal large language models via over-trust penalty and retrospection-allocation](https://md.research.elamur.ai/free/en/papers/huang-2023-opera-alleviating-hallucination-multi-modal/). In Proceedings of the IEEE/CVF Conference on Computer Vis…
13. Bo Jin, Nuno Gonçalves, Leandro Cruz, Iurii Medvedev, Yuanyu Yu, and Jiujiang Wang. 2024. Simulated multimodal deep facial diagnosis. Expert Systems with Applications, 252:123881. [find](https://scholar.google.com/scholar?q=Bo+Jin%2C+Nuno+Gon%C3%A7alves%2C+Leandro+Cruz%2C+Iurii+Medvedev%2C+Yuanyu+Yu%2C+and+Jiujiang+Wang%2E+2024%2E+Simulated+multimodal+deep+facial+diagnosis%2E+Expert+Systems+with+Applications%2C+252%3A123881%2E)
14. Yanwei Li, Yuechen Zhang, Chengyao Wang, Zhisheng Zhong, Yixin Chen, Ruihang Chu, Shaoteng Liu, and Jiaya Jia. 2024. [Mini-gemini: Mining the potential of multi-modality vision language models](https://md.research.elamur.ai/free/en/papers/li-2024-mini-gemini-mining-potential-multi/). arXiv preprint arXiv:2403.18814.
15. 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.
16. Haotian Liu, Chunyuan Li, Yuheng Li, Bo Li, Yuanhan Zhang, Sheng Shen, and Yong Jae Lee. 2024a. Llava-next: Improved reasoning, ocr, and world knowledge. [find](https://scholar.google.com/scholar?q=Haotian+Liu%2C+Chunyuan+Li%2C+Yuheng+Li%2C+Bo+Li%2C+Yuanhan+Zhang%2C+Sheng+Shen%2C+and+Yong+Jae+Lee%2E+2024a%2E+Llava%2Dnext%3A+Improved+reasoning%2C+ocr%2C+and+world+knowledge%2E)
17. Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. 2024b. [Visual instruction tuning](https://md.research.elamur.ai/free/en/papers/liu-2023-visual-instruction-tuning/). Advances in neural information processing systems, 36.
18. Pan Lu, Swaroop Mishra, Tanglin Xia, Liang Qiu, Kai-Wei Chang, Song-Chun Zhu, Oyvind Tafjord, Peter Clark, and Ashwin Kalyan. 2022. Learn to explain: Multimodal reasoning via thought chains for science question answering. Advances in Neural Information Processing Systems, 35:2507–2521. [find](https://scholar.google.com/scholar?q=Pan+Lu%2C+Swaroop+Mishra%2C+Tanglin+Xia%2C+Liang+Qiu%2C+Kai%2DWei+Chang%2C+Song%2DChun+Zhu%2C+Oyvind+Tafjord%2C+Peter+Clark%2C+and+Ashwin+Kalyan%2E+2022%2E+Learn+to+explain%3A+Multimodal+reasoning+via+thought+chains+for+scienc)
19. Daniel Omeiza, Skyler Speakman, Celia Cintas, and Komminist Weldermariam. 2019. Smooth grad-cam++: An enhanced inference level visualization technique for deep convolutional neural network models. arXiv preprint arXiv:1908.01224. — [arXiv:1908.01224](https://arxiv.org/abs/1908.01224) [find](https://scholar.google.com/scholar?q=Daniel+Omeiza%2C+Skyler+Speakman%2C+Celia+Cintas%2C+and+Komminist+Weldermariam%2E+2019%2E+Smooth+grad%2Dcam%2B%2B%3A+An+enhanced+inference+level+visualization+technique+for+deep+convolutional+neural+network+models%2E+arX)
20. 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.
21. Amanpreet Singh, Vivek Natarajan, Meet Shah, Yu Jiang, Xinlei Chen, Dhruv Batra, Devi Parikh, and Marcus Rohrbach. 2019. Towards vqa models that can read. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8317–8326. [find](https://scholar.google.com/scholar?q=Amanpreet+Singh%2C+Vivek+Natarajan%2C+Meet+Shah%2C+Yu+Jiang%2C+Xinlei+Chen%2C+Dhruv+Batra%2C+Devi+Parikh%2C+and+Marcus+Rohrbach%2E+2019%2E+Towards+vqa+models+that+can+read%2E+In+Proceedings+of+the+IEEE%2FCVF+conference+on+)
22. Eric Todd, Millicent L Li, Arnab Sen Sharma, Aaron Mueller, Byron C Wallace, and David Bau. 2023. [Function vectors in large language models](https://md.research.elamur.ai/free/en/papers/todd-2023-function-vectors-large-language-models/). arXiv preprint arXiv:2310.15213.
23. Lean Wang, Lei Li, Damai Dai, Deli Chen, Hao Zhou, Fandong Meng, Jie Zhou, and Xu Sun. 2023a. [Label words are anchors: An information flow perspective for understanding in-context learning](https://md.research.elamur.ai/free/en/papers/wang-2023-label-words-anchors-information-flow/). arXiv preprint arXiv:2305.14160.
24. Weihan Wang, Qingsong Lv, Wenmeng Yu, Wenyi Hong, Ji Qi, Yan Wang, Junhui Ji, Zhuoyi Yang, Lei Zhao, Song XiXuan, et al. 2023b. Cogvlm: Visual expert for large language models. [find](https://scholar.google.com/scholar?q=Weihan+Wang%2C+Qingsong+Lv%2C+Wenmeng+Yu%2C+Wenyi+Hong%2C+Ji+Qi%2C+Yan+Wang%2C+Junhui+Ji%2C+Zhuoyi+Yang%2C+Lei+Zhao%2C+Song+XiXuan%2C+et+al%2E+2023b%2E+Cogvlm%3A+Visual+expert+for+large+language+models%2E)
25. Jinfeng Wei and Xiaofeng Zhang. 2024. Dopra: Decoding over-accumulation penalization and re-allocation in specific weighting layer. Proceedings of the 32nd ACM International Conference on Multimedia. [find](https://scholar.google.com/scholar?q=Jinfeng+Wei+and+Xiaofeng+Zhang%2E+2024%2E+Dopra%3A+Decoding+over%2Daccumulation+penalization+and+re%2Dallocation+in+specific+weighting+layer%2E+Proceedings+of+the+32nd+ACM+International+Conference+on+Multimedia%2E)
26. Shengqiong Wu, Hao Fei, Leigang Qu, Wei Ji, and Tat-Seng Chua. 2023. [Next-gpt: Any-to-any multimodal llm](https://md.research.elamur.ai/free/en/papers/wu-2023-next-gpt-any-any-multimodal/). arXiv preprint arXiv:2309.05519.
27. Jiabo Ye, Anwen Hu, Haiyang Xu, Qinghao Ye, Ming Yan, Yuhao Dan, Chenlin Zhao, Guohai Xu, Chenliang Li, Junfeng Tian, et al. 2023. mplug-docowl: Modularized multimodal large language model for document understanding. arXiv preprint arXiv:2307.02499. — [arXiv:2307.02499](https://arxiv.org/abs/2307.02499) [find](https://scholar.google.com/scholar?q=Jiabo+Ye%2C+Anwen+Hu%2C+Haiyang+Xu%2C+Qinghao+Ye%2C+Ming+Yan%2C+Yuhao+Dan%2C+Chenlin+Zhao%2C+Guohai+Xu%2C+Chenliang+Li%2C+Junfeng+Tian%2C+et+al%2E+2023%2E+mplug%2Ddocowl%3A+Modularized+multimodal+large+language+model+for+documen)

*Updated 2 November 2026*