beyonduncertainty-a3a88636·1 events·first seen Aliases: BeyondUncertainty
A new arXiv preprint introduces BeyondUncertainty, a method that uses verbalized confidence estimates from black-box LLMs to decide whether to invoke retrieval in RAG pipelines. The system elicits a provisional answer and confidence score, then routes low-confidence queries to TF-IDF retrieval while returning high-confidence answers directly. Evaluated across 27,000 policy instances on six QA benchmarks, the method achieves 0.483 mean token-level F1 versus 0.467 for always-retrieve, while reducing retrieved passages by 20.4%, though the confidence probe adds 28.2% token overhead.