Elamur Research⌕
☰
home / papers / 2026 / AdaptEvolve: Improving Efficiency of Evolutionary AI Agents through Adaptive Model Selection

AdaptEvolve: Improving Efficiency of Evolutionary AI Agents through Adaptive Model Selection

Paper · 2026 · Pretam Ray, Pratik Prabhanjan Brahma, Zicheng Liu et al.

Summary

Evolutionary coding agents such as OpenEvolve call LLMs many times during search. They usually pick models through fixed weights or schedules, so expensive models get used even for easy refinement steps. AdaptEvolve replaces this with routing guided by uncertainty. A small model (Qwen3-4B) first proposes each mutation. Token-entropy confidence statistics from that generation then feed a lightweight classifier, which decides whether to keep the draft or escalate to a large model (Qwen3-32B). A 50-sample warm-up trains an initial shallow decision tree. Because the MAP-Elites archive keeps raising task difficulty over the run, a Hoeffding Adaptive Tree then updates the router online. The authors evaluated the method on LiveCodeBench v5 and MBPP. Averaged over these two benchmarks, it lowers compute by about 37.9% and keeps 97.5% of the accuracy of always using the 32B model. It also beats random routing and a standard cascade on efficiency. This suggests that a model's own confidence signals can make multi-LLM agent pipelines much cheaper without a heavy external router.

Key points

  • Confidence features: the router uses four sequence-level signals, all computed from top-k token log-probabilities:
    • Mean Confidence (MC)
    • Lowest Group Confidence (LGC)
    • Tail Confidence (TC)
    • Bottom-K% Confidence (BWC)
  • Router and cost setup: a Gini decision tree (max depth 5) is bootstrapped on N = 50 examples, then adapted online with a Hoeffding Adaptive Tree from the River package. Experiments ran on 8× AMD Instinct MI250 GPUs. Cost was normalized so that one 32B call = 1 unit and one 4B call = 0.125 units.
  • LiveCodeBench v5 (880 problems):
    • Accuracy was 73.6% vs. 75.2% for the 32B-only baseline, with 34.4% less compute.
    • Efficiency was 35.4, compared with 23.7 for 32B-only and 26.3 for a cascade.
    • Online adaptation added 2.4 accuracy points over the static tree.
  • MBPP (974 problems):
    • About 85% of calls stayed on the small model.
    • Accuracy was 91.3% vs. 94.0%, with 41.5% less cost.
    • Efficiency was 132.3 vs. 79.7.
  • Transfer and limitation: with LLaMA 3.1 8B/70B on HumanEval, the router reached 74.3% vs. 69.4% for random routing at the same ratio, a 1.56× speedup over 70B-only. The method needs access to logprobs, so it cannot be used with closed APIs that return only text.

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

🎓Indian Institute of Technology KharagpurKharagpur · India24 works · 714 citations

 Fields, topics and concepts

📚 Shelves 4The 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 works

Cite

Cite as Pretam Ray, Pratik Prabhanjan Brahma, Zicheng Liu and Emad Barsoum. 2026. AdaptEvolve: Improving Efficiency of Evolutionary AI Agents through Adaptive Model Selection. 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.2019

@inproceedings{ray2026adaptevolve,
  title = {{AdaptEvolve: Improving Efficiency of Evolutionary AI Agents through Adaptive Model Selection}},
  author = {Pretam Ray, Pratik Prabhanjan Brahma, Zicheng Liu and Emad Barsoum},
  booktitle = {Findings of the Association for Computational Linguistics: ACL 2026},
  year = {2026},
  doi = {10.18653/v1/2026.findings-acl.2019},
  url = {https://doi.org/10.18653/v1/2026.findings-acl.2019}
}

References

23 references, 9 of them with a card of ours. ↗ leads to the work, «find» searches for one we have no id for yet.

  1. Marah Abdin, Jyoti Aneja, Harkirat Behl, Sébastien Bubeck, Ronen Eldan, Suriya Gunasekar, Michael Harrison, Russell J Hewett, Mojan Javaheripi, Piero Kauffmann, and 1 others. 2024. Phi-4 technical report. arXiv preprint arXiv:2412.08905.
  2. Jacob Austin, Augustus Odena, Maxwell Nye, Maarten Bosma, Henryk Michalewski, David Dohan, Ellen Jiang, Carrie Cai, Michael Terry, Quoc Le, and 1 others. 2021. Program synthesis with large language models. arXiv preprint arXiv:2108.07732.
  3. Albert Bifet and Ricard Gavaldà. 2009. Adaptive learning from evolving data streams. pages 249–260. find
  4. Lars Buitinck, Gilles Louppe, Mathieu Blondel, Fabian Pedregosa, Andreas Mueller, Olivier Grisel, Vlad Niculae, Peter Prettenhofer, Alexandre Gramfort, Jaques Grobler, Robert Layton, Jake VanderPlas, Arnaud Joly, Brian Holt, and Gaël Varoquaux. 2013. API design for machine learning software: experie… find
  5. Lingjiao Chen, Matei Zaharia, and James Zou. 2023. Frugalgpt: How to use large language models while reducing cost and improving performance. arXiv preprint arXiv:2305.05176.
  6. Mark Chen, Jerry Tworek, Heewoo Jun, Qiming Yuan, Henrique Ponde de Oliveira Pinto, Jared Kaplan, Harri Edwards, Yuri Burda, Nicholas Joseph, Greg Brockman, Alex Ray, Raul Puri, Gretchen Krueger, Michael Petrov, Heidy Khlaaf, Girish Sastry, Pamela Mishkin, Brooke Chan, Scott Gray, and 39 others. 202… — Evaluating Large Language Models Trained on Code
  7. Yichao Fu, Xuewei Wang, Yuandong Tian, and Jiawei Zhao. 2025. Deep think with confidence. Preprint, arXiv:2508.15260. find
  8. Aaron Grattafiori, Abhimanyu Dubey, Abhinav Jauhri, Abhinav Pandey, Abhishek Kadian, Ahmad Al-Dahle, Aiesha Letman, Akhil Mathur, Alan Schelten, Alex Vaughan, and 1 others. 2024. The llama 3 herd of models. arXiv preprint arXiv:2407.21783.
  9. Naman Jain, King Han, Alex Gu, Wen-Ding Li, Fanjia Yan, Tianjun Zhang, Sida Wang, Armando Solar-Lezama, Koushik Sen, and Ion Stoica. 2024. Livecodebench: Holistic and contamination free evaluation of large language models for code. arXiv preprint arXiv:2403.07974.
  10. Sehoon Kim, Karttikeya Mangalam, Suhong Moon, Jitendra Malik, Michael W Mahoney, Amir Gholami, and Kurt Keutzer. 2023. Speculative decoding with big little decoder. Advances in Neural Information Processing Systems, 36:39236–39256. find
  11. Robert Tjarko Lange, Yuki Imajuku, and Edoardo Cetin. 2025. Shinkaevolve: Towards open-ended and sample-efficient program evolution. arXiv preprint arXiv:2509.19349. find
  12. Shudong Liu, Zhaocong Li, Xuebo Liu, Runzhe Zhan, Derek F. Wong, Lidia S. Chao, and Min Zhang. 2024. Can LLMs learn uncertainty on their own? expressing uncertainty effectively in a self-training manner. In Proceedings of the 2024 Conference on Empirical Methods in Natural Language Processing, pages… find
  13. Jacob Montiel, Max Halford, Saulo Martiello Mastelini, Geoffrey Bolmier, Raphael Sourty, Robin Vaysse, Adil Zouitine, Heitor Murilo Gomes, Jesse Read, Talel Abdessalem, and 1 others. 2021. River: machine learning for streaming data in python. find
  14. Jean-Baptiste Mouret and Jeff Clune. 2015. Illuminating search spaces by mapping elites. Preprint, arXiv:1504.04909. — [arXiv:1504.04909](https://arxiv.org/abs/1504.04909) find
  15. Alexander Novikov, Ngân Vũ, Marvin Eisenberger, Emilien Dupont, Po-Sen Huang, Adam Zsolt Wagner, Sergey Shirobokov, Borislav Kozlovskii, Francisco J. R. Ruiz, Abbas Mehrabian, M. Pawan Kumar, Abigail See, Swarat Chaudhuri, George Holland, Alex Davies, Sebastian Nowozin, Pushmeet Kohli, and Matej Ba… — AlphaEvolve: A coding agent for scientific and algorithmic discovery
  16. Isaac Ong, Amjad Almahairi, Vincent Wu, Wei-Lin Chiang, Tianhao Wu, Joseph E. Gonzalez, M Waleed Kadous, and Ion Stoica. 2024. Routellm: Learning to route llms with preference data. Preprint, arXiv:2406.18665. find
  17. Isaac Ong, Amjad Almahairi, Vincent Wu, Wei-Lin Chiang, Tianhao Wu, Joseph E. Gonzalez, M. Waleed Kadous, and Ion Stoica. 2025. RouteLLM: Learning to Route LLMs with Preference Data. arXiv preprint. ArXiv:2406.18665 [cs]. find
  18. Asankhaya Sharma. 2025. Openevolve: an open-source evolutionary coding agent. find
  19. Jianghui Wang, Vinay Joshi, Saptarshi Majumder, Xu Chao, Bin Ding, Ziqiong Liu, Pratik Prabhanjan Brahma, Dong Li, Zicheng Liu, and Emad Barsoum. 2025. Geak: Introducing triton kernel ai agent & evaluation benchmarks. Preprint, arXiv:2507.23194. find
  20. Xuezhi Wang, Jason Wei, Dale Schuurmans, Quoc Le, Ed Chi, Sharan Narang, Aakanksha Chowdhery, and Denny Zhou. 2022. Self-consistency improves chain of thought reasoning in language models. arXiv preprint arXiv:2203.11171.
  21. Jiaqi Wu, Qinlao Zhao, Zefeng Chen, Kai Qin, Yifei Zhao, Xueqian Wang, and Yuhang Yao. 2025. Gap: Graph-based agent planning with parallel tool use and reinforcement learning. Preprint, arXiv:2510.25320. — [arXiv:2510.25320](https://arxiv.org/abs/2510.25320) find
  22. 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.
  23. Zhaojian Yu, Kaiyue Feng, Yilun Zhao, Shilin He, Xiao-Ping Zhang, and Arman Cohan. 2025. Alpharesearch: Accelerating new algorithm discovery with language models. Preprint, arXiv:2511.08522. find

Updated

↑ back to top

This page in Markdown → https://md.research.elamur.ai/free/en/papers/ray-2026-adaptevolve-improving-efficiency-evolutionary-ai/