Can question features alone predict when to retrieve?

Adaptive retrieval aims to retrieve only when necessary, saving compute and reducing the misinformation RAG can introduce. But the dominant approach frames the decision as LLM-based uncertainty estimation — querying the model's internal states or outputs to gauge confidence — which is expensive and impractical at scale, often eating the very efficiency gains it promises. This paper proposes an LLM-independent alternative: decide whether to retrieve from lightweight external features of the question itself — 27 features across 7 groups, plus hybrids — with no model calls. Across 6 QA datasets it matches complex LLM-uncertainty methods on overall QA while delivering large efficiency gains, and outperforms them on complex questions.

The keeper is the relocation of the retrieval-trigger signal from the model's interior to the question's surface, and the demonstration that the two signal sources are complementary. This puts the paper in direct dialogue — and partial tension — with Can simple uncertainty estimates beat complex adaptive retrieval?, which argued uncertainty estimation beats heuristic triggers cheaply; the resolution is that "heuristics" here are not crude rules but learned external-feature predictors that rival uncertainty at lower cost and win on hard queries. It also sharpens the choice flagged by When should retrieval happen during model generation?: uncertainty-gating and external-feature-gating are alternative, combinable trigger sources, not a settled hierarchy.