Leveraging contextual confidence for smarter retrieval in large language models

Large Language Models (LLMs) often struggle with factual consistency in knowledge-intensive tasks due to limited internal knowledge. Retrieval-augmented generation (RAG) mitigates this by accessing external documents, yet static or indiscriminate retrieval can reduce efficiency and accuracy. We present SUGAR-L-Semantic Uncertainty Guided Adaptive Retrieval with Compression for Long Contexts-a lightweight, training-free framework that adaptively chooses between no, single-step, or multi-step retrieval based on entropy-derived confidence signals. SUGAR-L requires no dataset-specific supervision and leverages semantic entropy to measure epistemic uncertainty in the generation space. For multi-hop QA, it incorporates a plug-and-play compression module to handle lengthy retrieved contexts within model limits. Experiments across multiple QA benchmarks show that SUGAR-L improves answer quality while reducing redundant retrieval and computation. Ablation and sensitivity analyses further confirm its robustness, interpretability, and generalizability.