Paper accepted at ACL 2026

🎉 Happy to share that our paper “K-Merge: Online Continual Merging of Adapters for On-device Large Language Models” has been accepted at 𝐀𝐂𝐋 𝟐𝟎𝟐𝟔 (main, oral), which will be held in San Diego, California, this July! ☀️🏄

On-device deployment of Large Language Models frequently relies on Low-Rank Adapters (LoRAs) to support diverse downstream tasks under tight resource constraints. As users continuously request support for new tasks, new languages, new problem types, devices face a critical challenge: how to incorporate new adapters without forgetting previously learned capabilities, all within a limited storage budget.

In this work, we:

🎯 Introduce a new and practical setting for online continual merging of adapters in resource-constrained on-device LLMs under a storage budget.

🧠 Propose 𝐊-𝐌𝐞𝐫𝐠𝐞, a lightweight and data-free merging strategy that selects which adapters to merge by leveraging the history of past merges.

⚖️ Employ dynamic weighting to balance old and new capabilities during merging, avoiding catastrophic forgetting while staying within the compute and memory limits of real mobile devices.

📊 Conduct extensive experiments on real-world tasks representative of mobile device usage, showing consistent superiority over alternative merging strategies under realistic constraints.

Thanks to the amazing co-authors: Ondrej Bohdal, Taha Ceritli, Mete Ozay, Pietro Zanuttigh, and Umberto Michieli

📜 Preprint: https://arxiv.org/abs/2510.13537

🌐 Project Page: https://donaldssh.github.io/K-Merge/

💻 Code: https://github.com/donaldssh/K-Merge