The paper identifies why position-independent caching (PIC) for LLM agents degrades quality relative to full-context prefill. Through controlled experiments, it finds that the cached KV representations are faithful; the problem is in attention score computation, not the KV states themselves. Specifically, the model fails to select among multiple artifacts (discovery) rather than reading a single artifact (utilization).
Based on this insight, the authors propose Attuner, which inserts low-rank adapters into the query projections of the model. These adapters are trained via distillation from a full-prefill teacher to correct attention routing over frozen artifact caches. Attuner trains fewer than 0.05% of model parameters and requires no cache recomputation or full-context reference at inference.
On Qwen3-4B and Qwen3-8B across seven benchmarks covering skills, documents, memory, and code, Attuner substantially outperforms prior PIC methods and matches full-context prefill quality while providing up to 3.73x speedup in time to first token. This makes it attractive for agentic workloads with reusable content.