DynamiRAG: Dynamic Memory Retrieval-Augmented Generation with Entropy-Aware Hallucination Prediction

Retrieval-augmented generation (RAG) systems remain susceptible to hallucination when retrieved evidence is noisy, contradictory, or insufficient. While recent adaptive RAG systems such as Self-RAG and SmartRAG address retrieval decisions, they lack mechanisms to assess retrieval quality prior to generation or to detect contradictions within retrieved evidence sets.This paper introduces DynamiRAG, a framework that augments standard RAG pipelines with four novel components: (1) a Retrieval Entropy Hallucination Predictor (REHP) that formalizes retrieval uncertainty via Shannon entropy to estimate hallucination risk before generation; (2) an Evidence Contradiction Detector (ECD) that identifies inter-document conflicts through semantic-lexical divergence analysis; (3) Multi-Hop Memory Chains (MHMC) that perform multi-step reasoning through residual vector arithmetic without requiring knowledge graphs; and (4) a Short-Term Memory cache (STM) enabling session-persistent retrieval acceleration. We evaluate DynamiRAG on HotpotQA under fair corpus-retrieval conditions.

Our best configuration achieves 11.5% Exact Match and 16.8% F1, representing a +0.5% EM and +0.5% F1 improvement over the FiD baseline, with MHMCcontributing the largest individual gain (+2.0% EM over single-hop retrieval). We further report that Retrieval-Grounded Beam Scoring (RGBS), while theoretically motivated, decreases performance with small generators, providing insight into scale-dependent grounding effects. The ECD module successfully identifies contradictions in 97.5% of retrieval sets.

Our analysis demonstrates that entropy-based retrieval quality signals, contradiction-aware evidence filtering, and memory-augmented multi-hop reasoning each address distinct failure modes in RAG systems.