Fast LLM-Based Semantic Filtering: From a Unified Framework to an Adaptive Two-Phase Method

The article presents a novel adaptive two-phase method for semantic filtering using LLMs, addressing limitations of current cascade approaches. Key improvements include a hybrid model that integrates token-aware architectures, training proxies with oracle-derived soft labels, and adaptive composition of filtering strategies. Achieving a 90% accuracy target, the proposed method demonstrates a performance increase of 1.6-2.0x over previous techniques while maintaining high query success rates, highlighting significant efficiency gains for practitioners in LLM-based data processing.