CoinRAG: Contextualized Information Nugget KV Cache Reuse for Long-Context RAG
CoinRAG cuts RAG prefill work by reusing fine-grained “nugget” KV caches instead of whole chunks.
The paper says the method finds query-relevant semantic units inside retrieved chunks through two-stage retrieval, then assembles sliced KV representations with chunk-level context. Its target is long-context RAG where coarse chunk reuse still carries redundancy and noise. On LongBench multi-hop QA tasks, the authors report lower operational costs and a new Pareto frontier, with a 5.3% average relative F1 gain under a fast prefill latency budget. Source: ArXiv · AI/CL/LG's note.
The paper says the method finds query-relevant semantic units inside retrieved chunks through two-stage retrieval, then assembles sliced KV representations with chunk-level context. Its target is long-context RAG where coarse chunk reuse still carries redundancy and noise. On LongBench multi-hop QA tasks, the authors report lower operational costs and a new Pareto frontier, with a 5.3% average relative F1 gain under a fast prefill latency budget. Source: ArXiv · AI/CL/LG's note.
score 5