The paper "Breaking the Length Barrier: LLM-Enhanced CTR Prediction in Long Textual User Behaviors" addresses the efficiency bottleneck of large language models (LLMs) in processing long textual user behaviors for click-through rate (CTR) prediction. The authors propose Behavior Aggregated Hierarchical Encoding (BAHE), a novel hierarchical architecture that decouples the encoding of user behaviors from inter-behavior interactions. BAHE first extracts embeddings of atomic user behaviors using the LLM's pre-trained shallow layers, storing them in an offline database to avoid redundant encoding. The deeper, trainable layers of the LLM then facilitate intricate inter-behavior interactions, generating comprehensive user embeddings. This separation reduces computational complexity and enhances efficiency. Extensive experimental results show that BAHE reduces training time and memory by five times for CTR models using LLMs, especially with longer user sequences. The method has been successfully deployed in a real-world system, enabling daily updates of 50 million CTR data on 8 A100 GPUs, making LLMs practical for industrial CTR prediction. The paper also includes a detailed analysis of the method's complexity and empirical comparisons with baseline methods, demonstrating its effectiveness in both efficiency and performance.The paper "Breaking the Length Barrier: LLM-Enhanced CTR Prediction in Long Textual User Behaviors" addresses the efficiency bottleneck of large language models (LLMs) in processing long textual user behaviors for click-through rate (CTR) prediction. The authors propose Behavior Aggregated Hierarchical Encoding (BAHE), a novel hierarchical architecture that decouples the encoding of user behaviors from inter-behavior interactions. BAHE first extracts embeddings of atomic user behaviors using the LLM's pre-trained shallow layers, storing them in an offline database to avoid redundant encoding. The deeper, trainable layers of the LLM then facilitate intricate inter-behavior interactions, generating comprehensive user embeddings. This separation reduces computational complexity and enhances efficiency. Extensive experimental results show that BAHE reduces training time and memory by five times for CTR models using LLMs, especially with longer user sequences. The method has been successfully deployed in a real-world system, enabling daily updates of 50 million CTR data on 8 A100 GPUs, making LLMs practical for industrial CTR prediction. The paper also includes a detailed analysis of the method's complexity and empirical comparisons with baseline methods, demonstrating its effectiveness in both efficiency and performance.