HiLaR Framework Uses Hierarchical Latent Reasoning for LLM Recommendations

HiLaR optimizes LLM-based recommendations by constructing temporal-guided hierarchical user preference representations. It aligns broad preferences with fine-grained current intents through layer-aware reinforcement optimization to reduce the inference overhead of natural-language reasoning.