In real‑world scenarios, recommendation often occurs in a shared physical environment, requiring conversational recommenders to reason over dialogue history, co‑observed scenes, and in‑scene item attributes. This setting is called situated conversational recommendation (SCR) and poses two main challenges: accurately capturing user preferences that are grounded in the current context, and generating responses that satisfy both user needs and situational constraints.
Re2A formulates SCR as a structured "reason‑then‑align" process. It first introduces rubric‑based preference reasoning, where automated rubrics guide the model to produce explicit preference states. Based on these states, a preference‑conditioned optimization aligns response generation with two objectives: user preference satisfaction and situation consistency.
Extensive experiments on two SCR benchmarks show that Re2A consistently outperforms state‑of‑the‑art methods in precision and context awareness. The implementation is publicly available at https://github.com/DongdingLin/Re2A.
Review