Baidu’s RAG-LLM predicts content expiration for fresher search results

In commercial web search, traditional approaches rely on fixed time-window filters (e.g., “last 30 days”) to rank content freshness, but these often return chronologically recent yet semantically expired results. To solve this, researchers from Baidu (Tingyu Chen, Wenkai Zhang, Li Gao, et al.) propose a Query-Aware Dynamic Content Expiration Prediction Framework that uses large language models (LLMs) enhanced with retrieval-augmented generation (RAG). The framework extracts fine-grained temporal cues from each document, then uses an LLM to infer a query-specific “validity horizon”—a semantic boundary that defines when information becomes obsolete for a given user intent. This shifts timeliness from a static recency check to a dynamic reasoning task.

Deployed on live Baidu search traffic, the system incorporates hallucination mitigation strategies to ensure reliability. Offline and online A/B tests demonstrate significant improvements in search freshness and user experience metrics, validating LLM-driven reasoning for industrial-scale semantic expiration prediction. The paper, accepted at SIGIR 2026, shows that context-aware LLMs can outperform one-size-fits-all freshness heuristics. For professionals, this means search engines that return not just recent results, but *relevant* results that stay current only as long as the information actually holds value.