AutoAdapt enables efficient domain updates with modular adapters
AutoAdapt: Automatic Domain Discovery Enables Low-Cost Extensibility
Machine Learning
Summary
Instruction-tuned AI models often work with many different types of tasks or topics, which can change over time. The paper presents AutoAdapt, a way to add new topics or domains without retraining the entire model by training small separate modules called adapters. AutoAdapt figures out hidden domains on its own and trains these adapters independently, so they don’t interfere with each other. This approach matches the performance of training one big adapter for all domains, but is more flexible and cheaper to update.
What this means in practice
- •For natural language processing engineers: Add new language or task domains to AI models by training independent adapters without full retraining and avoiding interference.
- •For machine learning system maintainers: Update deployed AI systems with ongoing new data domains in parallel without disrupting existing specialized modules or full model retraining.
- •For enterprise ai platform teams: Build customizable AI services that dynamically incorporate new user domains with modular adapters, reducing update overhead and scaling easily.$Commercial implications: Enables selling adaptable AI products that efficiently incorporate client-specific domains with modular extension capabilities.
Authors
Josh McGiff, Salma Mekaoui, Robert Shanahan, Nikola S. Nikolov
Abstract
Instruction-tuned models are deployed into environments where domains are heterogeneous and evolve, yet adding new domains or data typically requires costly retraining. We present AutoAdapt, a modular framework that incorporates new domains and data via targeted single-adapter training without modifying other adapters. The framework automatically discovers latent domains, uses them to train per-domain Low-Rank Adaptation (LoRA) adapters independently in parallel and performs parameter-free routing. Across 14 domain-specific benchmarks and GPT-4o pairwise judgements, AutoAdapt achieves parity with a LoRA adapter trained on all domains without requiring full-model retraining. We also find evidence of specialisation effect convergence across independent discovery methods. Overall, training each adapter on its own domain prevents domain interference by construction, thus enabling modular, taxonomy-free domain specialisation without aggregate performance loss or full model retraining.