Adaptive RAG decides when to retrieve or correct evidence based on signals such as confidence, relevance, support and retrieval quality. In multi‑step retrieval these local signals must be merged into a persistent view of what the current evidence supports, what is missing, and what action should follow. Existing approaches often treat each signal as an independent trigger, which fragments the evidence state across the reasoning trajectory – a problem we call evidence‑state fragmentation. We therefore introduce BELIEFRAG, a closed‑loop controller that explicitly tracks sufficiency, reliability, conflict, uncertainty, evidence gaps and acquisition cost, and selects among retrieval, query rewriting, verification, answering, stopping and abstention. Across six QA benchmarks with gpt‑oss‑120b, BELIEFRAG achieves a mean token $F1$ of 0.572 using 3.89k tokens per question, outperforming fixed iterative retrieval (0.555 $F1$) while saving 39% of tokens. The same quality‑cost advantage transfers to Qwen3‑32B, reaching 0.552 versus 0.523 $F1$ and reducing tokens by 35%. Further analysis shows that most gains stem from corrective re‑retrieval rather than pruning alone, several belief dimensions are redundant, and calibrated answerability plays the strongest operational role. Calibration improves threshold stability across related evidence sources, though source shift can still invalidate the same decision signal.
Review: BELIEFRAG demonstrates a viable path for improving adaptive RAG efficiency through explicit state management, achieving lower token consumption without sacrificing answer quality, and offering insights for future multimodal or cross‑domain retrieval systems.