SituRecBench: A Benchmark for Situated Recommendation in 3D Interactive Environments

Recommender systems are transitioning from static retrieval engines into interactive assistants grounded in physical environments. However, current paradigms largely treat recommendation as a decoupled ranking task, neglecting the situated complexities of spatial navigation, visual grounding, and multi-turn social interaction. To bridge this gap, we introduce SituRecBench, a novel interactive benchmark designed to evaluate situated recommendation agents within high-fidelity 3D environments. SituRecBench reformulates the recommendation process as a closed-loop interaction between a recommendation assistant and a user with evolving latent preferences and realistic behaviors. Our benchmark features diverse real-world shopping scenarios (e.g., supermarkets and home furnishings stores), requiring assistants to integrate preference elicitation, spatial reasoning, and long-horizon conversation. We introduce a simulator-in-the-loop evaluation protocol along with a multi-dimensional metric taxonomy that quantifies communication fluency, recommendation utility, and navigation effectiveness. Through extensive evaluation of state-of-the-art vision-language models, we reveal critical challenges in situated recommendation, such as inefficient preference elicitation, coherent conversation transition, and poor spatial grounding. SituRecBench provides a rigorous, reproducible foundation for advancing the frontier of situated recommendation in interactive, visually-grounded environments.