Dynamic Knowledge Grounding for Retrieval-Augmented Generation: Uncertainty-Aware Evidence Fusion with Adaptive Retrieval Control

Despite the success of Retrieval-Augmented Generation (RAG) as a paradigm to provide grounding of LLM output in external knowledge, current RAG approaches have three core limitations: (1) uniform retrieval for every query, which wastes computation and adds irrelevant noise; (2) evidence fusion via shallow combination, failing to model differential relevance and reliability among retrieved passages; and (3) no explicit grounding verification to prevent hallucination from propagating despite retrieval augmentation. We present FORGE (Factually-Optimized Retrieval for Grounded gEneration), a principled framework that addresses all three limitations through three tightly integrated modules: the Adaptive Retrieval Controller (ARC), which learns an uncertainty-driven policy for when, what, and how much to retrieve; the Evidence Fusion Network (EFN), which performs multi-granularity cross-attention fusion with passage-level reliability weighting; and the Grounding Verification Module (GVM), which provides token-level factual consistency checking with calibrated confidence scores. We provide rigorous theoretical analysis establishing: (i) \(\:O(√TlogK)\) retrieval regret bounds for the adaptive policy; (ii) information-theoretic optimality guarantees for evidence fusion under the rate-distortion framework; (iii) PAC-Bayes generalization bounds for grounding verification; and (iv) end-to-end hallucination rate reduction guarantees. Extensive experiments on eight benchmarks (Natural Questions, TriviaQA, HotpotQA, FEVER, ASQA, ELI5, PopQA, TruthfulQA) demonstrate that FORGE achieves state-of-the-art performance with 2.1–4.8 point improvements over the best baselines while reducing hallucination rates by 38.7% and retrieval costs by 42.3%.