Elamur Research⌕
☰
home / papers / 2026 / Beyond End-to-End: Dynamic Chain Optimization for Private LLM Adaptation on the Edge

Beyond End-to-End: Dynamic Chain Optimization for Private LLM Adaptation on the Edge

Paper · 2026 · Yebo Wu, Jingguang Li, Chunlin Tian et al.

Summary

Federated fine-tuning lets LLMs be adapted to private data that never leaves user devices. However, even adapter-based approaches still require the whole model in memory, which excludes phones and other edge hardware with roughly 4–12GB. The authors show that frozen base weights, not activations or adapters, dominate this footprint. Their framework, CHAINFED, replaces end-to-end updates with a sequential chain: one layer's adapter is trained to convergence, frozen, and the next layer is then added, so only a small part of the network is active at once.

Three components address weaknesses of this layer-wise scheme:

  • Dynamic Layer Co-Tuning jointly trains a sliding window of neighbouring adapters to close semantic gaps.
  • Globally Perceptive Optimization blends the full-model objective into each local loss.
  • Function-Oriented Adaptive Tuning decides at which layer tuning should begin.

On text classification (e.g., YELP-P, AGNEWS) and instruction tuning with LLaMA models, CHAINFED beats existing methods by up to 46.46% average accuracy while cutting peak memory several-fold. This makes federated LLM adaptation feasible on edge devices.

Key points

  • Memory profiling with adapters: LLaMA2-7B needs about 27.1GB (base parameters 91.2%, activations 6.9%, adapter 1.9%), and LLaMA2-13B needs about 50.8GB (parameters 94.1%).
  • In BERT experiments, memory-constrained participation lowers accuracy on YELP-P by 8.5% (IID) and 11.8% (non-IID) compared with an idealized setting.
  • Tuning only upper layers (threshold T = 0.8) works better than tuning the full chain, by up to 4.43% on AGNEWS. On that dataset it also exceeds idealized Full Adapters by 3.29%, with 2.02× faster convergence and 3.47× less communication.
  • Instruction tuning (MMLU, BBH, DROP, CRASS):
    • LLaMA2-7B with Q = 6 cuts peak memory by 4.29× while improving on all four benchmarks.
    • LLaMA3.1-8B gains up to 10.71% average accuracy with 3.45× less memory.
  • Ablation on DistilBERT: removing DLCT, GPO, or FOAT lowers average accuracy by 11.64%, 13.73%, and 6.52%, respectively.

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

🎓Hong Kong University of Science and TechnologyHong Kong · China170 works · 26,341 citations🎓University of MacauMacao · China43 works · 2,368 citations🎓The Hong Kong University of Science and Technology (Guangzhou)Guangzhou · China67 works · 1,765 citations

 Fields, topics and concepts

📚 Shelves 4The field of science this work belongs to (OpenAlex fields).Computer Science · 18,614 worksArtificial Intelligence · 17,965 worksArts and Humanities · 17,450 worksLanguage and Linguistics · 17,296 works

🧭 Topics 1Research subjects and languages the work is about.Language models & pretraining · 3,894 works

Cite

Cite as Yebo Wu, Jingguang Li, Chunlin Tian, KaHou Tam, Zhijiang Guo and Li Li. 2026. Beyond End-to-End: Dynamic Chain Optimization for Private LLM Adaptation on the Edge. 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.839

@inproceedings{wu2026beyond,
  title = {{Beyond End-to-End: Dynamic Chain Optimization for Private LLM Adaptation on the Edge}},
  author = {Yebo Wu, Jingguang Li, Chunlin Tian, KaHou Tam, Zhijiang Guo and Li Li},
  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.839},
  url = {https://doi.org/10.18653/v1/2026.acl-long.839}
}

References

3 references, 1 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, Yunfei Chu, Zeyu Cui, Kai Dang, Xiaodong Deng, Yang Fan, Wenbin Ge, Yu Han, Fei Huang, et al. 2023. Qwen technical report. arXiv preprint arXiv:2309.16609.
  2. Jieming Bian, Yuanzhe Peng, Lei Wang, Yin Huang, and Jie Xu. 2025. A survey on parameter-efficient fine-tuning for foundation models in federated learning. arXiv preprint arXiv:2504.21099. — [arXiv:2504.21099](https://arxiv.org/abs/2504.21099) find
  3. Dongqi Cai, Yaozong Wu, Shangguang Wang, Felix Xiaozhu Lin, and Mengwei Xu. 2022. Fedadapter: Efficient federated learning for modern nlp. arXiv preprint arXiv:2205.10162. \#\# Ethical Considerations — [arXiv:2205.10162](https://arxiv.org/abs/2205.10162) find

Updated

This page in Markdown → https://md.research.elamur.ai/free/en/papers/wu-2026-beyond-end-end-dynamic-chain/