Agent4POI boosts POI recommendations by 23% with context-aware reasoning

Agent4POI tackles a fundamental limitation in multimodal point-of-interest recommendation: existing systems encode each POI once as a static embedding, failing to reason why a cafe works for solo work on Monday but not a group celebration on Friday. The authors prove that no pre-computed encoder can satisfy context-sensitive ranking under standard bilinear scoring, motivating inference-time item-side representation. Agent4POI inverts this computation using a four-phase LLM agent: it first generates context-specific affordance queries, then executes a five-step cross-modal chain-of-thought over image, review, and metadata evidence. The result is an uncertainty-aware affordance representation grounded in Gibsonian affordance theory, aligned with user preferences via semantic caching for low-latency ranking.

On three standard POI benchmarks and three evaluation configurations (standard, cold-start, context-shift), Agent4POI delivers a 23.2% relative gain over the strongest baseline. Under context-shift—where user scenarios change dramatically—it degrades only 7.5% compared to 16–17% for leading baselines. In cold-start scenarios, where no historical user data exists, Agent4POI outperforms the best content-based baseline by up to 2.4x, while ID-based methods fail entirely. The framework's ability to dynamically adapt POI representations to user context marks a significant advance for recommendation systems in dynamic environments like travel, daily routines, or event planning.