Technical Evolution of OPPO's Conversational AI Assistant XiaoBu

The article introduces OPPO's conversational AI assistant XiaoBu, outlining its technical evolution and the reflections and practices of the development team.

It is structured into four parts: a brief history of AI assistants, an overview of the XiaoBu product, a four‑year evolution review with team insights, and a forward‑looking outlook.

The historical review traces AI assistants from MIT's rule‑based system in 1966, through Apple's Siri in 2011 and Microsoft's XiaoIce in 2014, to recent large‑scale pretrained dialogue models that have achieved near‑human performance in specific scenarios.

XiaoBu was launched in 2019 for OPPO smartphones and later expanded to watches, TVs, and other IoT devices; it now serves over 300 million users, with monthly active users exceeding 1.4 billion and more than 30 billion interactions per month.

The evolution strategy is described in three user‑need layers—efficiency, intelligence, and emotional connection—and details skill construction (300+ skills covering 2500 intents) while highlighting challenges such as semantic understanding, open‑domain QA, and heavy reliance on manual labeling.

Anti‑failure (anti‑idiot) measures include model upgrades with multi‑task decoding, knowledge‑enhanced pretraining, cost‑effective inference via unified representations, and automated defect discovery and optimization, such as noise rejection and semi‑automatic error mining.

Understanding users is addressed through modeling user attributes, behaviors, and emotions, implementing memory for personalized interactions, leveraging feedback loops, and building a reputation system to improve satisfaction.

Future directions emphasize richer device integration, multimodal digital humans, persona development, and emotional engines to create more trustworthy and engaging AI assistants.

The Q&A segment discusses multi‑task instruction handling, distinguishing task‑oriented from chit‑chat intents, fine‑tuning retrieval models, reducing dependence on labeled data, and strategies for coreference resolution across dialogue turns.

The presentation concludes with thanks to the audience.