InduceKV Enables Fixed-Footprint Continual Adaptation for Multimodal LLMs

Multimodal large language models (MLLMs) face significant challenges in continual adaptation, particularly when operating under strict memory constraints. Traditional methods often lead to an accumulating "adaptation state" through repeated parameter updates or expanding replay stores, which is unsustainable for long-term deployment. This research addresses the problem of fixed-footprint continual adaptation, where the deployed memory budget for adaptation must remain constant, and the core backbone model is kept unchanged. The proposed solution is InduceKV, a novel retrieval-based method. InduceKV works by storing carefully selected training prefixes as "attention-ready" memory entries. Each entry comprises a frozen retrieval key and compact layer-wise key-value (KV) payloads that can be dynamically appended to the model's self-attention cache. A bilevel selection process ensures that under a strict memory budget, the chosen memory balances current-task likelihood, retention of old knowledge, and coverage in the retrieval space. Experiments show InduceKV consistently outperforms other methods like PEFT and replay under matched memory budgets across various continual learning scenarios.