Elamur Research⌕
☰
home / papers / 2026 / Forest Before Trees: Latent Superposition for Efficient Visual Reasoning

Forest Before Trees: Latent Superposition for Efficient Visual Reasoning

Paper · 2026 · Yubo Wang, Juntian Zhang, Yichen Wu et al.

Summary

Chain-of-Thought helps vision-language models reason step by step. However, spelling out rationales as discrete text loses much of the continuous visual signal. Existing latent-reasoning methods avoid text but still train each hidden state to predict one exact next token, which pushes the model to commit to specific objects too early.

Laser addresses this with Dynamic Windowed Alignment Learning (DWAL). Each latent step is trained to match a soft distribution over the entire remaining part of the reasoning chain. This window shrinks as reasoning proceeds, so the model first holds broad scene-level concepts and only later narrows to details.

The training data are "cognitive scanpaths": global-to-local sequences of concept tokens produced by GPT-4o, with no bounding-box supervision. Two mechanisms keep training stable:

  • Self-Refined Superposition uses the model's own window estimate as a soft target.
  • Entropy-Regularized Intervention falls back to ground-truth tokens when the model's uncertainty is high.

On six benchmarks, Laser outperforms the latent baseline Monet while using far fewer tokens. Its latent steps also remain readable through the LM head, which makes efficient, inspectable visual reasoning practical.

Key points

  • On average across six benchmarks (MMStar, MMVP, BLINK, HallusionBench, SEED-Bench2-Plus, HRBench), Laser beats Monet by 5.03% and reduces inference tokens by more than 97%. In the paper's illustrative example, a CoT answer takes 105 tokens versus 15 for Laser.
  • The scanpath prompts follow Navon's Global Precedence Hypothesis. The setup is weakly supervised and has no Region-of-Interest annotations.
  • Hidden states can be projected through the frozen LM head to show the top-k tokens at each step. An MMStar case traces the model's reasoning from the audience, to spatial cues, to the final answer option.
  • In the ablations:
    • Replacing DWAL with plain next-token prediction mainly hurts fine-grained perception tasks.
    • Using a fixed window instead of a dynamic one mainly hurts complex reasoning tasks.
  • The method was tested only on static, standard-resolution images. Video and high-resolution documents remain untested, and there is no metric yet for judging the quality of latent intermediate steps.

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

🎓Fudan UniversityShanghai · China167 works · 20,987 citations🎓Harvard UniversityCambridge · United States112 works · 42,023 citations🎓Renmin University of ChinaBeijing · China114 works · 16,118 citations

 Fields, topics and concepts

📚 Shelves 2The field of science this work belongs to (OpenAlex fields).Computer Science · 16,667 worksComputer Vision and Pattern Recognition · 2,869 works

🧭 Topics 1Research subjects and languages the work is about.Reasoning · 1,426 works

Cite

Cite as Yubo Wang, Juntian Zhang, Yichen Wu, Yankai Lin, Nils Lukas and Yuhan Liu. 2026. Forest Before Trees: Latent Superposition for Efficient Visual Reasoning. 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. DOI 10.18653/v1/2026.acl-long.517

@inproceedings{wang2026forest,
  title = {{Forest Before Trees: Latent Superposition for Efficient Visual Reasoning}},
  author = {Yubo Wang, Juntian Zhang, Yichen Wu, Yankai Lin, Nils Lukas and Yuhan Liu},
  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.517},
  url = {https://doi.org/10.18653/v1/2026.acl-long.517}
}

References

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

  1. Shuai Bai, Yuxuan Cai, Ruizhe Chen, Keqin Chen, Xionghui Chen, Zesen Cheng, Lianghao Deng, Wei Ding, Chang Gao, Chunjiang Ge, Wenbin Ge, Zhifang Guo, Qidong Huang, Jie Huang, Fei Huang, Binyuan Hui, Shutong Jiang, Zhaohai Li, Mingsheng Li, and 45 others. 2025a. Qwen3-vl technical report. Preprint, a…
  2. Shuai Bai, Keqin Chen, Xuejing Liu, Jialin Wang, Wenbin Ge, Sibo Song, Kai Dang, Peng Wang, Shijie Wang, Jun Tang, Humen Zhong, Yuanzhi Zhu, Mingkun Yang, Zhaohai Li, Jianqiang Wan, Pengfei Wang, Wei Ding, Zheren Fu, Yiheng Xu, and 8 others. 2025b. Qwen2.5-vl technical report. Preprint, arXiv:2502.1…
  3. Ali Behrouz, Peilin Zhong, and Vahab Mirrokni. 2024. Titans: Learning to memorize at test time. Preprint, arXiv:2501.00663. find
  4. Shibo Hao, Sainbayar Sukhbaatar, DiJia Su, Xian Li, Zhiting Hu, Jason Weston, and Yuandong Tian. 2025. Training large language models to reason in a continuous latent space. Preprint, arXiv:2412.06769. find
  5. Wenxuan Huang, Bohan Jia, Zijie Zhai, Shaosheng Cao, Zheyu Ye, Fei Zhao, Zhe Xu, Yao Hu, and Shaohui Lin. 2025. Vision-r1: Incentivizing reasoning capability in multimodal large language models. arXiv preprint arXiv:2503.06749. find
  6. Bangzheng Li, Ximeng Sun, Jiang Liu, Ze Wang, Jialian Wu, Xiaodong Yu, Hao Chen, Emad Barsoum, Muhao Chen, and Zicheng Liu. 2025. Latent visual reasoning. arXiv preprint arXiv:2509.24251. — [arXiv:2509.24251](https://arxiv.org/abs/2509.24251) find
  7. Bo Li, Yuanhan Zhang, Dong Guo, Renrui Zhang, Feng Li, Hao Zhang, Kaichen Zhang, Peiyuan Zhang, Yanwei Li, Ziwei Liu, and Chunyuan Li. 2024a. Llava-onevision: Easy visual task transfer. Preprint, arXiv:2408.03326.
  8. Chao Chen, Zhixin Ma, Yongqi Li, Yupeng Hu, Yinwei Wei, Wenjie Li, and Liqiang Nie. 2025. Reasoning in the dark: Interleaved vision-text reasoning in latent space. Preprint, arXiv:2510.12603. find
  9. Bohao Li, Yuying Ge, Yixiao Ge, Guangzhi Wang, Rui Wang, Ruimao Zhang, and Ying Shan. 2024b. SEED-Bench-2-Plus: Benchmarking multimodal large language models with text-rich visual comprehension. arXiv preprint arXiv:2404.16790. — [arXiv:2404.16790](https://arxiv.org/abs/2404.16790) find
  10. Lin Chen, Jinsong Li, Xiaoyi Dong, Pan Zhang, Yuhang Zang, Zehui Chen, Haodong Duan, Jiaqi Wang, Yu Qiao, Dahua Lin, and Feng Zhao. 2024a. Are we on the right way for evaluating large vision-language models? In Advances in Neural Information Processing Systems, volume 37, pages 27056–27087. Curran A…
  11. Chengzhi Liu, Yuzhe Yang, Yue Fan, Qingyue Wei, Sheng Liu, and Xin Eric Wang. 2025. Reasoning within the mind: Dynamic multimodal interleaving in latent space. Preprint, arXiv:2512.12623. — [arXiv:2512.12623](https://arxiv.org/abs/2512.12623) find
  12. Zhe Chen, Jiannan Wu, Wenhai Wang, Weijie Su, Guo Chen, Sen Xing, Muyan Zhong, Qinglong Zhang, Xizhou Zhu, Lewei Lu, Bin Li, Ping Luo, Tong Lu, Yu Qiao, and Jifeng Dai. 2024b. Internvl: Scaling up vision foundation models and aligning for generic visual-linguistic tasks. In Proceedings of the IEEE/C…
  13. Shuai Dong, Siyuan Wang, Xingyu Liu, and Zhongyu Wei. 2025. Interleaved latent visual reasoning with selective perceptual modeling. Preprint, arXiv:2512.05665.
  14. Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. 2023. Visual instruction tuning. In Advances in Neural Information Processing Systems, volume 36, pages 34892–34916. Curran Associates, Inc.
  15. Jizheng Ma, Xiaofei Zhou, Geyuan Zhang, Yanlong Song, and Han Yan. 2026. Multimodal reasoning via latent refocusing. Preprint, arXiv:2511.02360. — [arXiv:2511.02360](https://arxiv.org/abs/2511.02360) find
  16. David Navon. 1977. Forest before trees: The precedence of global features in visual perception. Cognitive psychology, 9(3):353–383. find
  17. OpenAI, Josh Achiam, Steven Adler, Sandhini Agarwal, Lama Ahmad, Ilge Akkaya, Florencia Leoni Aleman, Diogo Almeida, Janko Altenschmidt, Sam Altman, Shyamal Anadkat, Red Avila, Igor Babuschkin, Suchir Balaji, Valerie Balcom, Paul Baltescu, Haiming Bao, Mohammad Bavarian, Jeff Belgum, and 262 others.… — GPT-4 Technical Report
  18. Tan-Hanh Pham and Chris Ngo. 2025. Multimodal chain of continuous thought for latent-space reasoning in vision-language models. Preprint, arXiv:2508.12587. — [arXiv:2508.12587](https://arxiv.org/abs/2508.12587) find
  19. Arijit Ray, Ahmed Abdelkader, Chengzhi Mao, Bryan A. Plummer, Kate Saenko, Ranjay Krishna, Leonidas Guibas, and Wen-Sheng Chu. 2025. Mull-tokens: Modality-agnostic latent thinking. Preprint, arXiv:2512.10941. find
  20. Hao Shao, Shengju Qian, Han Xiao, Guanglu Song, Zhuofan Zong, Letian Wang, Yu Liu, and Hongsheng Li. 2024. Visual cot: Advancing multi-modal language models with a comprehensive dataset and benchmark for chain-of-thought reasoning. In Advances in Neural Information Processing Systems, volume 37, pag… find
  21. Guohao Sun, Hang Hua, Jian Wang, Jiebo Luo, Sohail Dianat, Majid Rabbani, Raghuveer Rao, and Zhiqiang Tao. 2025. Latent chain-of-thought for visual reasoning. Preprint, arXiv:2510.23925. — [arXiv:2510.23925](https://arxiv.org/abs/2510.23925) find
  22. Jintao Tong, Jiaqi Gu, Yujing Lou, Lubin Fan, Yixiong Zou, Yue Wu, Jieping Ye, and Ruixuan Li. 2025. Sketch-in-latents: Eliciting unified reasoning in mllms. Preprint, arXiv:2512.16584. — [arXiv:2512.16584](https://arxiv.org/abs/2512.16584) find
  23. Shengbang Tong, Zhuang Liu, Yuexiang Zhai, Yi Ma, Yann LeCun, and Saining Xie. 2024. Eyes wide shut? exploring the visual shortcomings of multimodal llms. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 9568–9578.
  24. Haozhe Wang, Chao Qu, Zuming Huang, Wei Chu, Fangzhen Lin, and Wenhu Chen. 2025a. Vl-rethinker: Incentivizing self-reflection of vision-language models with reinforcement learning. Preprint, arXiv:2504.08837. find
  25. Qixun Wang, Yang Shi, Yifei Wang, Yuanxing Zhang, Pengfei Wan, Kun Gai, Xianghua Ying, and Yisen Wang. 2025b. Monet: Reasoning in latent visual space beyond images and language. Preprint, arXiv:2511.21395. — [arXiv:2511.21395](https://arxiv.org/abs/2511.21395) find
  26. Weiyun Wang, Zhangwei Gao, Lixin Gu, Hengjun Pu, Long Cui, Xingguang Wei, Zhaoyang Liu, Linglin Jing, Shenglong Ye, Jie Shao, Zhaokai Wang, Zhe Chen, Hongjie Zhang, Ganlin Yang, Haomin Wang, Qi Wei, Jinhui Yin, Wenhao Li, Erfei Cui, and 56 others. 2025c. Internvl3.5: Advancing open-source multimodal…
  27. Wenbin Wang, Liang Ding, Minyan Zeng, Xiabin Zhou, Li Shen, Yong Luo, Wei Yu, and Dacheng Tao. 2025d. Divide, conquer and combine: A training-free framework for high-resolution image perception in multimodal large language models. Proceedings of the AAAI Conference on Artificial Intelligence, 39(8):… find
  28. Zhenhailong Wang, Xuehang Guo, Sofia Stoica, Haiyang Xu, Hongru Wang, Hyeonjeong Ha, Xiusi Chen, Yangyi Chen, Ming Yan, Fei Huang, and Heng Ji. 2026. Perception-aware policy optimization for multimodal reasoning. Preprint, arXiv:2507.06448. — [arXiv:2507.06448](https://arxiv.org/abs/2507.06448) find
  29. Jason Wei, Xuezhi Wang, Dale Schuurmans, Maarten Bosma, brian ichter, Fei Xia, Ed Chi, Quoc V Le, and Denny Zhou. 2022. Chain-of-thought prompting elicits reasoning in large language models. In Advances in Neural Information Processing Systems, volume 35, pages 24824–24837. Curran Associates, Inc.
  30. Yige Xu, Xu Guo, Zhiwei Zeng, and Chunyan Miao. 2025. Softcot: Soft chain-of-thought for efficient reasoning with llms. Preprint, arXiv:2502.12134.
  31. Zeyuan Yang, Xueyang Yu, Delin Chen, Maohao Shen, and Chuang Gan. 2025. Machine mental imagery: Empower multimodal reasoning with latent visual tokens. Preprint, arXiv:2506.17218. — [arXiv:2506.17218](https://arxiv.org/abs/2506.17218) find
  32. Eric Zelikman, Georges Harik, Yijia Shao, Varuna Jayasiri, Nick Haber, and Noah D. Goodman. 2024. Quiet-star: Language models can teach themselves to think before speaking. Preprint, arXiv:2403.09629.
  33. Huanyu Zhang, Wenshan Wu, Chengzu Li, Ning Shang, Yan Xia, Yangyu Huang, Yifan Zhang, Li Dong, Zhang Zhang, Liang Wang, Tieniu Tan, and Furu Wei. 2025a. Latent sketchpad: Sketching visual thoughts to elicit multimodal reasoning in mllms. Preprint, arXiv:2510.24514. — [arXiv:2510.24514](https://arxiv.org/abs/2510.24514) find
  34. Juntian Zhang, Chuanqi Cheng, Yuhan Liu, Wei Liu, Jian Luan, and Rui Yan. 2025b. Weaving context across images: Improving vision-language models through focus-centric visual chains. In Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pa… find
  35. Juntian Zhang, Song Jin, Chuanqi Cheng, Yuhan Liu, Yankai Lin, Xun Zhang, Yufei Zhang, Fei Jiang, Guojun Yin, Wei Lin, and Rui Yan. 2025c. Viper: Empowering the self-evolution of visual perception abilities in vision-language model. Preprint, arXiv:2510.24285. — [arXiv:2510.24285](https://arxiv.org/abs/2510.24285) find
  36. Ziwei Zheng, Michael Yang, Jack Hong, Chenxiao Zhao, Guohai Xu, Le Yang, Chao Shen, and Xing Yu. 2026. Deepeyes: Incentivizing "thinking with images" via reinforcement learning. Preprint, arXiv:2505.14362. — [arXiv:2505.14362](https://arxiv.org/abs/2505.14362) find
  37. Deyao Zhu, Jun Chen, Xiaoqian Shen, Xiang Li, and Mohamed Elhoseiny. 2023. Minigpt-4: Enhancing vision-language understanding with advanced large language models. Preprint, arXiv:2304.10592.

Updated

↑ back to top

This page in Markdown → https://md.research.elamur.ai/free/en/papers/wang-2026-forest-before-trees-latent-superposition/