Decoupling Knowledge and Task Subspaces for Composable Parametric Retrieval Augmented Generation

The core contribution is a decomposition method that forces document adapters to encode factual content and task behavior into separate, non-overlapping subspaces, making the composition step at inference a cleaner arithmetic operation rather than a collision between entangled representations. The practical bet here is that clean subspace separation is achievable at training time without sacrificing per-adapter retrieval quality.

This connects to the long-context serving work covered in 'Unifying Sparse Attention with Hierarchical Memory for Scalable Long-Context LLM Serving' from the same day. Both papers are attacking the same underlying pressure: production systems need to handle more information sources simultaneously, and the current architectures buckle under that load. The difference is that SPIN works at the inference-serving layer while this work intervenes at the adapter-training layer. Together they suggest the field is converging on a view that multi-document reasoning requires coordinated fixes at multiple levels of the stack, not a single architectural swap.

The falsifiable test is whether teams building multi-adapter RAG pipelines can reproduce the composition stability gains on domain-shift benchmarks outside the paper's evaluation set. If third-party replication holds across at least two heterogeneous domain pairs within the next six months, the subspace decomposition approach will likely get absorbed into standard adapter training recipes.