Think-to-Personalize framework uses LLM reasoning to boost e-commerce search by 0.46%

Dense retrieval is the backbone of e-commerce search, but standard models are query-centric and ignore the reasoning capabilities of modern LLMs. In a new paper accepted at CIKM 2026, researchers from multiple institutions introduce Think-to-Personalize (TTP), a framework that explicitly reasons over a user's historical purchase sequence to deduce latent personalized needs. TTP generates an intent-enhanced query, which is then encoded into a unified dense embedding alongside the original query and user context.

The architecture employs a two-stage training paradigm: supervised fine-tuning (SFT) establishes cold-start capabilities so the model works without user history, then reinforcement learning (RL) using Group Relative Policy Optimization (GRPO) aligns the reasoning process with actual retrieval utility. This overcomes the limitations of implicit embedding interactions used by existing personalized retrieval methods, which fail to disambiguate intent from noisy historical behavior. Extensive experiments on proprietary and public benchmarks show TTP significantly outperforms state-of-the-art baselines. Crucially, online A/B tests demonstrated a +0.46% lift in order volume, proving practical business impact. TTP establishes a new paradigm for reasoning-driven personalized dense retrieval in local-lifestyle e-commerce.