Think2Go framework boosts next-POI prediction with LLM reasoning and RL

Next Point-of-Interest (POI) recommendation aims to predict a user's next destination from historical check-ins, but existing non-reasoning models struggle with sparse, complex mobility data and shallow contextual features. Large language models (LLMs) also underperform when semantic IDs (SIDs) are trained separately, lacking deep understanding. To fix this, researchers propose Think2Go, a generative framework that unifies supervised fine-tuning (SFT) and reinforcement learning (RL)-based reasoning in a single architecture. This joint optimization enables both memorization of user behavior patterns and adaptive exploration of diverse preferences, leveraging test-time computational scaling to uncover richer spatial-temporal patterns.

Think2Go introduces two advantage weighting mechanisms to calibrate policy optimization. First, prompt epistemic uncertainty is estimated via kernel density methods, measuring alignment between query and user history to encourage exploration when uncertainty is high. Second, reward-informed advantage scaling normalizes rewards against their maximum, adjusting update magnitudes for stability and reducing overfitting to noisy signals. Together, these form an implicit curriculum learning strategy that delivers instance-aware policy updates and prevents entropy collapse. Accepted to KDD 2026 as an oral presentation, Think2Go demonstrates a promising path toward more intelligent, reasoning-driven location recommendations.