How Vivo’s Blue Heart XiaoV Leverages LLMs to Transform Conversational Recommendations

In 2023 Vivo launched the Blue Heart XiaoV intelligent assistant and upgraded it with DeepSeek‑R1 in June 2025, enhancing AI capabilities for personalized recommendation.

DataFun: What is the core technical challenge of the "question‑after‑answer" recommendation scenario compared with traditional feed or e‑commerce recommendation? Liang Tianan: The key challenge is generating high‑quality recommendation items in real time, as the system must create candidate items from fragmented, instantaneous user intents within the dialogue context.

The generated items directly affect downstream recommendation performance, and evaluating them requires multiple dimensions such as relevance, usefulness, diversity and safety, which differ from traditional click‑through or conversion metrics.

DataFun: Could you describe the original recall and ranking modules and their bottlenecks? Liang Tianan: Initially, XiaoV used a single LLM to generate recommendations, leading to unstable quality, lack of diversity, and no internal ranking mechanism.

To address this, a multi‑stage pipeline with recall, ranking and re‑ranking was introduced.

Multiple parallel LLM recall paths generate diverse candidate items by injecting different roles, tasks and constraints into prompts.

LLM‑based ranking predicts CTR by fine‑tuning on exposure‑click data, replacing traditional statistical ranking models.

Offline evaluation employs a fine‑tuned reward model (LLM) to score generated items, improving evaluation efficiency over rule‑based methods.

DataFun: How are LLMs combined with traditional modules? Liang Tianan: LLMs are used in recall to generate rich candidates, in ranking to estimate CTR, and as a reward model for offline evaluation. Prompt engineering techniques such as role injection, few‑shot safety prompts, and chain‑of‑thought reasoning improve relevance and safety.

DataFun: How do you balance recommendation quality with mobile resource constraints? Liang Tianan: Model quantization to INT8, prefix‑cache and KV‑cache (including CPU‑based KV‑cache) are applied to reduce latency and memory usage while maintaining performance.

DataFun: What are the biggest remaining technical obstacles? Liang Tianan: (1) Incomplete evaluation systems – current reward‑model scores still differ from online metrics, prompting research into hybrid offline metrics combining CTR and generative rewards. (2) Balancing diversity and relevance in dialogue‑driven recommendations.

DataFun: What is the future potential of multimodal LLMs on mobile? Liang Tianan: Multimodal models can align visual, sensor and textual signals to infer user intent, enabling richer services such as travel assistance or "one‑click screen" queries. Key preparations include cross‑modal alignment, lightweight knowledge distillation, and on‑device private knowledge bases.