Innovation Series: Advanced Science (ISSN 2938-9933, CNKI Indexed)

DPLite-CRS: Towards Dynamic and Efficient Conversational Recommendation

Yongsen Wei

School of Physics and Information Engineering, Guilin Normal College, Guilin 541199, China

Abstract: Conversational recommender systems must incorporate new preferences while retaining relevant dialogue history and controlling response cost. These requirements motivate a joint design of preference tracking and computation allocation. This vision paper proposes DPLite-CRS, a framework combining incremental preference updates, parameter-efficient adaptation, and adaptive recommendation routing. A compact state is up-dated from each new dialogue turn; low-rank adapters support domain adaptation; and a drift-aware controller selects between direct item retrieval and a more expensive contextual reranking path. The central hypothesis is that computation should depend on the evidence needed at each turn rather than follow a fixed processing path. We outline an illustrative formulation, identify failure cases, and propose an evaluation protocol covering ranking quality, preference revision, latency, and memory. The framework remains a research proposal: its effectiveness and deployment benefits require empirical validation.

Keywords: Terms—conversational recommendation; Preference tracking; Parameter-efficient adaptation; Adaptive inference; Vision paper

References