2608.05161

Total: 1

#1 SemiAdapt-Instruct: Extensible Instruction Tuning via Latent Domain-Specialised Adapters [PDF] [Copy] [Kimi] [REL]

Authors: Josh McGiff, Salma Mekaoui, Robert Shanahan, Nikola S. Nikolov

Instruction-tuned LLMs are deployed into environments where domains evolve, yet extending a fine-tuned model's capabilities without full retraining remains an unsolved practical challenge. We present SemiAdapt-Instruct, a modular framework that discovers latent instruction domains, trains per-domain LoRA adapters in parallel, and performs parameter-free routing, incorporating new domains via single-adapter training without modifying existing components. SemiAdapt-Instruct outperforms full model fine-tuning across all configurations on both ROUGE-L and LLM-as-a-judge evaluation, while matching single LoRA fine-tuning and delivering extensibility that monolithic approaches cannot provide. We empirically demonstrate this extensibility by showing that updating a single adapter with new domain data outperforms all monolithic baselines. Our study also finds that independent discovery methods converge on the same specialisation-friendly domains. These findings demonstrate that decomposing heterogeneous instruction data into latent domains enables extensible NLP systems where evolving domains require only targeted single-adapter updates, eliminating the need for full model retraining.

Subjects: Computation and Language , Machine Learning

Publish: 2026-05-26 07:01:35 UTC