Elamur Research⌕
☰
home / papers / 2026 / MAXS: Meta-Adaptive Exploration with LLM Agents

MAXS: Meta-Adaptive Exploration with LLM Agents

Paper · 2026 · Jian Zhang, Zhiyuan Wang, Zhangqi Wang et al.

Summary

The paper addresses inference-time decision making in tool-using LLM agents, where step-by-step methods such as CoT and ToT decide greedily without looking ahead, and small early mistakes can push a multi-tool trajectory off course. MCTS reduces this instability but consumes far more tokens. The authors propose MAXS, an inference-time framework that combines search and code tools with planning. At each step it rolls candidate continuations a few steps forward and scores them with three signals: an advantage term for the gain in foresight probability, a step-level variance term inspired by Lyapunov stability, and a slope-level variance term for consistency across steps. A trajectory convergence rule stops rollouts once the candidate paths agree, which saves compute. MAXS was tested with MiMo-VL-7B, Qwen2.5-VL-7B and Qwen2.5-VL-32B on MathVista, OlympiadBench, EMMA, TheoremQA and MATH. It beat CoT, ToT, MCTS, Guided Decoding and ϕ-Decoding on accuracy at a fraction of MCTS-level token cost. This makes foresighted agent reasoning more practical. Code is available at github.com/exoskeletonzj/MAXS.

Key points

  • MiMo-VL-7B-SFT: MAXS reaches a 63.46% average versus 57.04% for ToT, the strongest baseline. On MathVista it scores 85.5% versus 77.2% for CoT.
  • Other backbones:
    • Qwen2.5-VL-7B-Instruct: 42.85% average versus 35.42% for the best baseline (Guided Decoding).
    • Qwen2.5-VL-32B-Instruct on EMMA: 39.33% versus 33.00% for ϕ-Decoding.
  • Efficiency: According to the authors, MCTS needs roughly 1000× more tokens for similar accuracy. A 4-step lookahead is the sweet spot, since going to 6 steps adds about 49.8% more tokens (2.05×10⁷ → 3.07×10⁷) with no meaningful accuracy gain.
  • Tool ablation on MiMo:
    • The full model's 63.46% falls to 60.81% without code and 56.36% without search.
    • Removing both tools gives 52.07%.
    • On MathVista, removing code alone drops accuracy from 85.5% to 73.0%.
  • Value weights and step cap:
    • The best weighting is α=0.3 for step variance and β=0.2 for slope variance, giving 63.5% versus 55.2% for advantage-only scoring.
    • Most problems finish within 4–8 steps, which supports the 13-step cap.

Our summary of the paper, in our own words — not its text. The work belongs to its authors; read it at the links above.

 Fields, topics and concepts

📚 Shelves 5The field of science this work belongs to (OpenAlex fields).Computer Science · 18,614 worksArtificial Intelligence · 17,965 worksInformation Systems · 1,730 worksDecision Sciences · 249 worksManagement Science and Operations Research · 134 works

Cite

Cite as Jian Zhang, Zhiyuan Wang, Zhangqi Wang, Yu He, Haoran Luo, li Yuan, Lingling Zhang, Rui Mao, Qika Lin and Jun Liu. 2026. MAXS: Meta-Adaptive Exploration with LLM Agents. 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.670

@inproceedings{zhang2026maxs,
  title = {{MAXS: Meta-Adaptive Exploration with LLM Agents}},
  author = {Jian Zhang, Zhiyuan Wang, Zhangqi Wang, Yu He, Haoran Luo, li Yuan, Lingling Zhang, Rui Mao, Qika Lin and Jun Liu},
  booktitle = {Findings of the Association for Computational Linguistics: ACL 2026},
  year = {2026},
  doi = {10.18653/v1/2026.findings-acl.670},
  url = {https://doi.org/10.18653/v1/2026.findings-acl.670}
}

References

44 references, 17 of them with a card of ours. ↗ leads to the work, «find» searches for one we have no id for yet.

  1. Bowen Baker, Ingmar Kanitscheider, Todor Markov, Yi Wu, Glenn Powell, Bob McGrew, and Igor Mordatch. 2019. Emergent tool use from multi-agent autocurricula. In International conference on learning representations. find
  2. EN Barron and H Ishii. 1989. The bellman equation for minimizing the maximum cost. Nonlinear Anal. Theory Methods Applic., 13(9):1067–1090. find
  3. Mark Chen, Jerry Tworek, Heewoo Jun, Qiming Yuan, Henrique Ponde De Oliveira Pinto, Jared Kaplan, Harri Edwards, Yuri Burda, Nicholas Joseph, Greg Brockman, et al. 2021. Evaluating large language models trained on code. arXiv preprint arXiv:2107.03374.
  4. Wenhu Chen, Ming Yin, Max Ku, Pan Lu, Yixin Wan, Xueguang Ma, Jianyu Xu, Xinyi Wang, and Tony Xia. 2023. Theoremqa: A theorem-driven question answering dataset. arXiv preprint arXiv:2305.12524.
  5. Yanxi Chen, Xuchen Pan, Yaliang Li, Bolin Ding, and Jingren Zhou. 2024. Ee-llm: Large-scale training and inference of early-exit large language models with 3d parallelism. In International Conference on Machine Learning, pages 7163–7189. PMLR. find
  6. Zhihui Chen, Kai He, Qingyuan Lei, Bin Pu, Jian Zhang, Yuling Xu, and Mengling Feng. 2026. Medforge: Interpretable medical deepfake detection via forgery-aware reasoning. arXiv preprint arXiv:2603.18577. find
  7. Wonje Choi, Woo Kyung Kim, Minjong Yoo, and Honguk Woo. 2024. Embodied cot distillation from llm to off-the-shelf agents. In Proceedings of the 41st International Conference on Machine Learning, pages 8702–8721. find
  8. Yumeng Fu, Jiayin Zhu, Lingling Zhang, Bo Zhao, Shaoxuan Ma, Yushun Zhang, Yanrui Wu, and Wenjun Wu. 2026. Geolaux: A benchmark for evaluating mllms’ geometry performance on long-step problems requiring auxiliary lines. The 64rd Annual Meeting of the Association for Computational Linguistics. find
  9. Bingzheng Gan, Yufan Zhao, Tianyi Zhang, Jing Huang, Li Yusu, Shu Xian Teo, Changwang Zhang, and Wei Shi. 2025. Master: A multi-agent system with llm specialized mcts. In Proceedings of the 2025 Conference of the Nations of the Americas Chapter of the Association for Computational Linguistics: Human… find
  10. Lin Gui, Cristina Gârbacea, and Victor Veitch. 2024. Bonbon alignment for large language models and the sweetness of best-of-n sampling. Advances in Neural Information Processing Systems, 37:2851–2885. find
  11. Fatemeh Haji, Mazal Bethany, Maryam Tabar, Jason Chiang, Anthony Rios, and Peyman Najafirad. 2024. Improving llm reasoning with multi-agent tree-of-thought validator agent. arXiv preprint arXiv:2409.11527. find
  12. Yunzhuo Hao, Jiawei Gu, Huichen Will Wang, Linjie Li, Zhengyuan Yang, Lijuan Wang, and Yu Cheng. 2025. Can mllms reason in multimodality? emma: An enhanced multimodal reasoning benchmark. arXiv preprint arXiv:2501.05444. — [arXiv:2501.05444](https://arxiv.org/abs/2501.05444) find
  13. Chaoqun He, Renjie Luo, Yuzhuo Bai, Shengding Hu, Zhen Leng Thai, Junhao Shen, Jinyi Hu, Xu Han, Yujie Huang, Yuxiang Zhang, et al. 2024. Olympiadbench: A challenging benchmark for promoting agi with olympiad-level bilingual multimodal scientific problems. arXiv preprint arXiv:2402.14008.
  14. Juha Heinonen. 2005. Lectures on Lipschitz analysis. 100. University of Jyväskylä. find
  15. Dan Hendrycks, Collin Burns, Saurav Kadavath, Akul Arora, Steven Basart, Eric Tang, Dawn Song, and Jacob Steinhardt. 2021. Measuring mathematical problem solving with the math dataset. arXiv preprint arXiv:2103.03874.
  16. Xu Huang, Weiwen Liu, Xiaolong Chen, Xingmei Wang, Hao Wang, Defu Lian, Yasheng Wang, Ruiming Tang, and Enhong Chen. 2024. Understanding the planning of llm agents: A survey. arXiv preprint arXiv:2402.02716.
  17. Bowen Jin, Hansi Zeng, Zhenrui Yue, Jinsung Yoon, Sercan Arik, Dong Wang, Hamed Zamani, and Jiawei Han. 2025. Search-r1: Training llms to reason and leverage search engines with reinforcement learning. arXiv preprint arXiv:2503.09516.
  18. 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. In Proceedings of the 29th symposium on operating systems principles, pages 611–626.
  19. Yifei Li, Weidong Guo, Lingling Zhang, Rongman Xu, Muye Huang, Hui Liu, Lijiao Xu, Yu Xu, and Jun Liu. 2026. Locomo-plus: Beyond-factual cognitive memory evaluation framework for llm agents. arXiv preprint arXiv:2602.10715. find
  20. Wenjin Liu, Haoran Luo, Xin Feng, Xiang Ji, Lijuan Zhou, Rui Mao, Jiapu Wang, Shirui Pan, and Erik Cambria. 2025. Lexgenius: An expert-level benchmark for large language models in legal general intelligence. arXiv preprint arXiv:2512.04578.
  21. Alejandro Torreno, Eva Onaindia, Antonín Komenda, and Michal Štolba. 2017. Cooperative multi-agent planning: A survey. ACM Computing Surveys (CSUR), 50(6):1–32. find
  22. Xingyao Wang, Yangyi Chen, Lifan Yuan, Yizhe Zhang, Yunzhu Li, Hao Peng, and Heng Ji. 2024. Executable code actions elicit better llm agents. In Forty-first International Conference on Machine Learning.
  23. Jason Wei, Xuezhi Wang, Dale Schuurmans, Maarten Bosma, Fei Xia, Ed Chi, Quoc V Le, Denny Zhou, et al. 2022. Chain-of-thought prompting elicits reasoning in large language models. Advances in neural information processing systems, 35:24824–24837.
  24. Yuxi Xie, Kenji Kawaguchi, Yiran Zhao, James Xu Zhao, Min-Yen Kan, Junxian He, and Michael Xie. 2023. Self-evaluation guided beam search for reasoning. In Advances in Neural Information Processing Systems, volume 36, pages 41618–41650. Curran Associates, Inc.
  25. Fangzhi Xu, Hang Yan, Chang Ma, Haiteng Zhao, Jun Liu, Qika Lin, and Zhiyong Wu. 2025a. ϕ-decoding: Adaptive foresight sampling for balanced inference-time exploration and exploitation. In Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers)… find
  26. Wenjin Liu, Haoran Luo, Xueyuan Lin, Haoming Liu, Tiesunlong Shen, Jiapu Wang, Rui Mao, and Erik Cambria. 2026. Prompt-r1: Collaborative automatic prompting framework via end-to-end reinforcement learning. arXiv preprint arXiv:2511.01016.
  27. Fangzhi Xu, Hang Yan, Qiushi Sun, Jinyang Wu, Zixian Huang, Muye Huang, Jingyang Gong, Zichen Ding, Kanzhi Cheng, Yian Wang, et al. 2026. Odysseyarena: Benchmarking large language models for long-horizon, active and inductive interactions. arXiv preprint arXiv:2602.05843. — [arXiv:2602.05843](https://arxiv.org/abs/2602.05843) find
  28. Pan Lu, Hritik Bansal, Tony Xia, Jiacheng Liu, Chunyuan Li, Hannaneh Hajishirzi, Hao Cheng, Kai-Wei Chang, Michel Galley, and Jianfeng Gao. 2023. Mathvista: Evaluating mathematical reasoning of foundation models in visual contexts. arXiv preprint arXiv:2310.02255.
  29. Jin Xu, Zhifang Guo, Jinzheng He, Hangrui Hu, Ting He, Shuai Bai, Keqin Chen, Jialin Wang, Yang Fan, Kai Dang, et al. 2025b. Qwen2. 5-omni technical report. arXiv preprint arXiv:2503.20215. — Qwen2.5-Omni Technical Report
  30. Haoran Luo, Yikai Guo, Qika Lin, Xiaobao Wu, Xinyu Mu, Wenhao Liu, Meina Song, Yifan Zhu, Luu Anh Tuan, et al. 2025. Kbqa-o1: Agentic knowledge base question answering with monte carlo tree search. arXiv preprint arXiv:2501.18922. — [arXiv:2501.18922](https://arxiv.org/abs/2501.18922) find
  31. Guangtao Nie, Rong Zhi, Xiaofan Yan, Yufan Du, Xiangyang Zhang, Jianwei Chen, Mi Zhou, Hongshen Chen, Tianhao Li, Ziguang Cheng, et al. 2024. A hybrid multi-agent conversational recommender system with llm and search engine in e-commerce. In Proceedings of the 18th ACM Conference on Recommender Syst… find
  32. Matthew Renze and Erhan Guven. 2024. Self-reflection in llm agents: Effects on problem-solving performance. arXiv preprint arXiv:2405.06682.
  33. Daniel Shevitz and Brad Paden. 2002. Lyapunov stability theory of nonsmooth systems. IEEE Transactions on automatic control, 39(9):1910–1914. find
  34. Hang Yan, Xinyu Che, Fangzhi Xu, Qiushi Sun, Zichen Ding, Kanzhi Cheng, Jian Zhang, Tao Qin, Jun Liu, and Qika Lin. 2026. Tide: Trajectory-based diagnostic evaluation of test-time improvement in llm agents. arXiv preprint arXiv:2602.02196. — [arXiv:2602.02196](https://arxiv.org/abs/2602.02196) find
  35. Hang Yan, Fangzhi Xu, Rongman Xu, Yifei Li, Jian Zhang, Haoran Luo, Xiaobao Wu, Luu Anh Tuan, Haiteng Zhao, Qika Lin, et al. 2025. Mur: Momentum uncertainty guided reasoning for large language models. arXiv preprint arXiv:2507.14958.
  36. Hang Yang, Hao Chen, Hui Guo, Yineng Chen, Ching-Sheng Lin, Shu Hu, Jinrong Hu, Xi Wu, and Xin Wang. 2024. Llm-medqa: Enhancing medical question answering through case studies in large language models. arXiv preprint arXiv:2501.05464. — [arXiv:2501.05464](https://arxiv.org/abs/2501.05464) find
  37. Shunyu Yao, Dian Yu, Jeffrey Zhao, Izhak Shafran, Tom Griffiths, Yuan Cao, and Karthik Narasimhan. 2023. Tree of thoughts: Deliberate problem solving with large language models. Advances in neural information processing systems, 36:11809–11822. find
  38. Li Yuan, Yi Cai, Xudong Shen, Qing Li, Qingbao Huang, Zikun Deng, and Tao Wang. 2025. Collaborative multi-lora experts with achievement-based multi-tasks loss for unified multimodal information extraction. In Proceedings of the Thirty-Fourth International Joint Conference on Artificial Intelligence,… find
  39. Li Yuan, Qingfei Huang, Bingshan Zhu, Yi Cai, Qingbao Huang, Changmeng Zheng, Zikun Deng, and Tao Wang. 2026. Hybrid-dmkg: A hybrid reasoning framework over dynamic multimodal knowledge graphs for multimodal multihop qa with knowledge editing. In Proceedings of the AAAI Conference on Artificial Inte… find
  40. Zihao Yue, Zhenru Lin, Yifan Song, Weikun Wang, Shuhuai Ren, Shuhao Gu, Shicheng Li, Peidian Li, Liang Zhao, Lei Li, et al. 2025. Mimo-vl technical report. CoRR.
  41. Jian Zhang, Shihao Qi, Yuxuan Dong, Li Yuan, Tiesunlong Shen, Weiping Fu, Bifan Wei, Haiping Zhu, and Jun Liu. 2025. Gkg-llm: A unified framework for generalized knowledge graph construction. Information Fusion, page 103956. find
  42. Jian Zhang, Zhangqi Wang, Haiping Zhu, Kangda Cheng, Kai He, Bo Li, Qika Lin, Jun Liu, and Erik Cambria. 2026a. Mars: Multi-agent adaptive reasoning with socratic guidance for automated prompt optimization. Proceedings of the AAAI Conference on Artificial Intelligence, 40(19):16307–16315. find
  43. Jian Zhang, Zhiyuan Wang, Zhangqi Wang, Fangzhi Xu, Qika Lin, Lingling Zhang, Rui Mao, Erik Cambria, and Jun Liu. 2026b. Maps: Multi-agent personality shaping for collaborative reasoning. Proceedings of the AAAI Conference on Artificial Intelligence, 40(19):16316–16324. find
  44. Zeyu Zhang, Quanyu Dai, Xiaohe Bo, Chen Ma, Rui Li, Xu Chen, Jieming Zhu, Zhenhua Dong, and Ji-Rong Wen. 2024. A survey on the memory mechanism of large language model based agents. ACM Transactions on Information Systems. find

Updated

↑ back to top

This page in Markdown → https://md.research.elamur.ai/free/en/papers/zhang-2026-maxs-meta-adaptive-exploration-llm/