LRT–ReviChain: Lightweight Orchestration for Improved Recency in Generative AI Services

Generative AI services remain unreliable for time-sensitive queries because retrieved information can rapidly become stale, citation–evidence alignment may degrade during generation, and tool pipelines are often brittle under real-world volatility. To address these limitations, we propose Lightweight Recency Theory (LRT) and its application-layer orchestration framework, LRT–ReviChain, which improve freshness and citation–evidence alignment without retraining foundation models or relying on a heavyweight retrieval backend. The proposed framework integrates four components: (i) recency-aware query routing and ranking, (ii) a unified URL Reader/ QA (Question Answering) module that converts heterogeneous web, document, and multimodal inputs into normalized textual evidence, (iii) an Evidence Pack that enforces strict sentence-level citation binding, and (iv) a self-check planner that verifies format compliance, freshness, span-level evidence support, and domain diversity before release. We evaluate LRT–ReviChain on four benchmark tracks: Web-Fresh QA (≤60 days), open-domain long-form QA, document/infographic QA, and video-grounded QA. Experimental results show consistent improvements in freshness and citation–evidence alignment, together with reductions in evidence-related errors, across all evaluation settings. On the Web-Fresh QA Core subset, the proposed method achieves RH = 0.98, CMatch@5 = 0.98, and EER = 0.02. We further validate portability on an open-source pipeline with five hosted model backends and across twelve public generative AI services. Ablation results indicate that recency control is the primary contributor to freshness improvement, while citation constraints and bounded repair play important roles in improving citation–evidence alignment and overall system reliability.