K-CARE: Knowledge-driven Symmetrical Contextual Anchoring and Analogical Prototype Reasoning for E-commerce Relevance

A team of researchers (Chen Yifei, Tian Zhixing, Wang Chenyang, and Cheng Ziguang) has introduced K-CARE, a novel framework for improving e-commerce search relevance by grounding LLM reasoning in external knowledge. The paper, submitted to arXiv on April 28, 2026, addresses a persistent problem: LLMs often fail on 'corner cases' like idiosyncratic queries or niche products due to knowledge boundaries in their parametric memory. Unlike existing approaches that optimize reasoning trajectories via reinforcement learning, K-CARE targets the root cause by extending the model's cognitive reach.

K-CARE comprises two synergistic components: Symmetrical Contextual Anchoring (SCA), which fills contextual voids by integrating implicit knowledge from user behavior data, and Analogical Prototype Reasoning (APR), which uses expert-curated prototypes to calibrate decision boundaries through in-context analogy. Extensive offline evaluations and online A/B tests on a major e-commerce platform showed K-CARE significantly outperforming state-of-the-art baselines, delivering substantial commercial impact by resolving knowledge-intensive relevance challenges. This approach offers a practical solution for deploying LLMs in real-world search systems where domain expertise is critical.