Large Language Model-Powered Query-Driven Event Timeline Summarization in Industrial Search

Researchers at Baidu have deployed a production system called QDET that uses large language models to build query-specific event timelines, according to a paper posted to arXiv on 26 May 2026 [1]. The system is designed to help search users understand how events evolve by organizing sub-events from millions of daily-retrieved documents [1]. The system, named Query-Driven Event Timeline Summarization, departs from traditional topic-centric timeline methods that aim for broad coverage. Instead, QDET identifies and organizes only those sub-events closely relevant to a user’s search query from noisy candidate sets [1]. The paper describes two technical innovations. The first is a multi-task supervised fine-tuning procedure that trains the model on three auxiliary tasks: temporal ordering, causal judgment, and timeline completion [1]. The second is a reinforcement learning-based summarization stage that enforces strict length constraints while preserving semantic quality. This stage achieved 88.2% length compliance and outperformed 671-billion-parameter models by 7.7 points on constraint satisfaction [1]. The fine-tuned model uses 7 billion parameters and reached a 76.2% F1 score on timeline summarization, slightly above the 76.1% zero-shot F1 score of the much larger DeepSeek-R1-671B model [1]. The authors note that QDET uses roughly 1% of the parameters of the 671-billion-parameter model, demonstrating that domain-specific optimization can yield comparable quality at a fraction of the computational cost [1]. Online A/B testing on Baidu Search validated the system’s real-world impact. Compared to single-task baselines, QDET delivered a 5.5% improvement in click-through rate, a 4.6% increase in dwell time, and a 4.4% gain in deeper exploration of search results [1]. The paper also reports that the timeline understanding learned by QDET transfers to a heat prediction task, indicating effective knowledge transfer to downstream applications [1]. The work arrives as large language models continue to reshape information retrieval. While symbolic AI methods dominated from the mid-1950s through the 1990s, neural network approaches reemerged strongly around 2012, when researchers used GPU power to scale up deep learning for vision, speech, and language tasks [4]. QDET’s architecture reflects this shift, applying neural fine-tuning and reinforcement learning to a specialized search problem rather than relying on hand-crafted rules or knowledge bases [1][4].