Knowledge Offloading: Decomposing LLMs into Sparse Backbones and Memory Modules
For LLM practitioners, this work proposes a method to separate general computation from specialized knowledge, enabling more modular and efficient model deployment.
Knowledge Offloading (KOFF) decomposes a pretrained LLM into a sparse shared backbone and domain-specific memory modules (LoRA adapters and learned KV caches). At 12% global sparsity, KOFF preserves most of the unpruned model's performance across Llama and Qwen models (3B-8B), while pruning without memories degrades sharply.
LLMs encode both general capabilities and domain-specific knowledge in a single set of parameters. We ask whether this capacity can be reorganized: keeping broadly useful computation in a shared backbone, while moving specialized knowledge into external memory modules. We propose \emph{knowledge offloading} (KOFF), a framework for decomposing a pretrained LLM into a sparse shared backbone and domain-specific memories. Starting from a frozen base model, we jointly learn a structured pruning mask and lightweight recovery modules, implemented as LoRA adapters and learned key-value caches. Across Llama and Qwen models from 3B to 8B, we find that non-trivial capacity can be moved out of the shared backbone without a large loss in model ability. At around 12\% global sparsity, KOFF preserves much of the unpruned model's performance, while pruning the same frozen model without memories degrades sharply. Ablations show that LoRA and learned KV memories are complementary, and specialization analyses suggest that the learned decomposition is meaningful: language-specific neurons are preferentially removed while language-general neurons largely remain in the backbone. These results suggest that knowledge can be reallocated between a shared core and swappable external memories.