Route, Reuse, Repurpose: Continual Adaptation of LLMs with Bounded Adapter Pools

Re-training Large Language Models (LLMs) each time a new task or domain emerges is neither practical nor cost-effective, making continual adaptation a central challenge. We introduce CaLLM, featuring metric-based meta-learning and modular parameter isolation via lightweight adapters with controllable plasticity. Each input is projected into a task-agnostic embedding space and routed to its nearest-prototype adapter through a non-parametric, task- and regime-agnostic router. New adapters are spawned when the signal is novel, or the least-used adapter is repurposed once capacity is full, yielding a bounded-cost form of intentional knowledge removal. CaLLM scales to many tasks by projecting inputs into a task-agnostic space and routing them to the nearest prototype adapter with a non-parametric, training-free router. As a result, training FLOPs, memory use, and routing overhead stay strictly constant regardless of the adapter pool size. We evaluate CaLLM on a cross-domain shift and a long-horizon dialogue stream, under both batched and online scenarios, where CaLLM outperforms baselines overall with the largest gains on the hardest tasks.