A new preprint on arXiv tackles a bottleneck in LLM-augmented data pipelines: multiple models often operate over the same long context—schemas, tool specifications, interaction history, and intermediate results—and each model recomputes its own attention. The authors propose reusing attention maps across models to reduce this redundant work.

The abstract frames the problem around heterogeneous models sharing substantial context. Prefix caching is the existing technique the paper starts from, but the authors argue it is not enough on its own; attention map reuse is presented as a way to accelerate inference across models that do not share weights.

Because this is a single source, there is no independent result to compare against. The claims are the authors' own, and the preprint has not been peer-reviewed.