TL;DR
Baidu researchers have developed a RAG-enhanced Large Language Model (LLM) framework to dynamically predict when web content expires based on specific user intent. By moving away from rigid, static time-window filters and inferring a query-specific "validity horizon," this framework successfully filters out semantically stale results while preserving long-term authoritative content at scale.
What Problem Does This Paper Address?
In commercial web search, aligning content freshness with what a user actually wants is a notorious challenge. Traditionally, industrial search engines rely on a "one-size-fits-all" approach: they apply a global "recency boost" to documents published within a fixed window (like the last 30 days) or indiscriminately penalize older content.
However, this creates a severe timestamp-validity mismatch. As shown in Figure 1, a universal expiration threshold simply does not exist. For a breaking news query like "Hong Kong Fire," a report from two days ago might already be obsolete if the fire has been extinguished and the situation has evolved. Conversely, for a query like "Traffic Regulations," a policy document published several years ago remains entirely valid. The authors address this problem by treating timeliness not as a static metadata attribute, but as a dynamic semantic relationship between query intent and content timelines.
Methods
The authors introduce a novel, LLM-based Query-Aware Dynamic Content Expiration Prediction Framework integrated directly into Baidu's production search engine. The architecture consists of four primary modules:
- Temporal Information Extraction Module: To prevent the LLM from getting bogged down by long-form text noise, this module distills raw documents into a concise "Focused Semantic Chunk Set" ($S_{focus}$). It maps out temporal expressions and scores text chunks based on keyword relevance and temporal alignment.
- Structured Prompting for Temporal Inference: The framework sets a strict absolute time anchor (the search time) and prompts the LLM to deduce a query-specific expiration threshold ($t^*_{exp}$). The model is trained using a joint optimization objective ($L_{temp}$) that penalizes logical contradictions and temporal granularity mismatches.
- Temporal Consistency Verification Module: To mitigate LLM hallucinations, the authors implement a dual-verification check. First, Consistency Forward-Backward Reasoning forces the LLM to generate a candidate date and then stress-test its own logic backward against the retrieved evidence. Second, Authority-Weighted Evidence Fusion resolves external conflicts when different web sources report divergent dates.
- Online Feature Application: Running heavy LLMs directly during live search queries is too slow. To meet strict online latency constraints, the framework distills the LLM's semantic judgments into a lightweight binary flag (1 for fresh, 0 for expired). This flag is injected into Baidu's core "Aurora" ranking model as a learnable soft constraint. A failsafe circuit breaker is also included to gracefully degrade to baseline ranking if a timeout or cache miss occurs.
Datasets
The framework was evaluated using a massive dataset constructed from real-world Baidu Search logs collected over a three-month period. This dataset comprised 1 million valid queries and 10 million candidate documents spanning diverse web domains.
Key Insights and Ideas
- Timeliness is Semantic, Not Chronological: The core realization is that a document's age is secondary to its semantic validity relative to the user's intent.
- Bridging the LLM Latency Gap: Instead of forcing a massive model to rank documents in real time, the framework cleverly decouples the heavy temporal reasoning (done via offline caching or distilled student models) from the live ranking loop using simple binary flags.
- Self-Correction via Forward-Backward CoT: By making the LLM double-check its own inferred expiration dates against the source text, the system drastically reduces the temporal drift and hallucinations that typically plague generative models.
Key Results
The framework was validated through extensive offline testing, human evaluation, and a 14-day live online A/B test on production traffic:
- Offline Metrics: The experimental model (Aurora-Expiry) achieved a +0.52% improvement in global Satisfaction Score and a substantial surge in Positive Navigation Rate (PNR)—up to +7.36% for highly time-sensitive queries.
- Online Production Impact: In the live A/B test, the framework drove a 12.81% drop in the median age of top results (
day_away@4) for High-Freshness queries. Crucially, this temporal cleanup translated directly to business value, yielding a +0.78% increase in Satisfactory Consumption and a +0.41% bump in Click-Through Rate (CTR). - Human Evaluation: Blind Good/Same/Bad (GSB) testing revealed an outstanding 12:2 Good-to-Bad ratio for complex, long-tail queries, confirming that the framework safely processes general traffic without destabilizing standard search results.
Limitations and Future Work
The authors do not explicitly dedicate a section to limitations, but a few engineering challenges are apparent. The system relies heavily on an offline dictionary cache and a distilled student LLM to maintain low latency; if a query triggers a cache miss or real-time inference times out, the system falls back to the baseline model, temporarily losing the benefits of dynamic expiration.
Future work will likely focus on improving the extraction of intrinsic content times when explicit publication dates are missing or ambiguous, as well as further optimizing the student model's distillation process to minimize reliance on the fallback circuit breaker.
Summary updated