Shengwang CEO Zhao Bin Proposes Five-Level Conversational AI Framework at iRTE2026

In his keynote, “Conversational AGI: The Next Chapter of Real-Time Interaction,” Zhao said conversational AI should not be judged only by response speed or natural-sounding language. He said it must understand a speaker’s intent, situation and relationships, express itself appropriately in complex settings and build long-term understanding.

The framework divides conversational capability into five levels: L1, expression from one’s own perspective, centered on answering a question and passive turn-taking; L2, expression based on interaction patterns, moving from passive response to active expression; L3, communication based on individual cognition, understanding identity, intent and situation; L4, communication based on group cognition, understanding social relationships and role differences in multi-person settings; L5, communication based on deep relationships, using long-term context and relationship changes to build deeper understanding.

Zhao said the five levels describe technical progress and a shift in the relationship between AI and people: from completing a question-and-answer round to understanding a person, then relationships among people, and eventually communication across time and situations. He said conversational AI is moving from passive turn-taking to more active, natural full-duplex dialogue and is beginning to understand personality, situation, emotion and relationships. Future systems still need to handle role division and communication order in complex social settings and understand long-term personal context.

Zhao said adoption of WebRTC by companies including OpenAI over the past year has brought more attention to real-time interaction technology in conversational AI. As AI moves from single-turn Q&A to continuous interaction, real-time infrastructure is becoming a link among models, agent runtimes and users, he said. Conversational scenarios place stricter demands on latency, stability and two-way interaction: AI must receive input, understand context and organize a response, handle interruptions and call tools to advance tasks. Noticeable delay in any part can break the naturalness of a conversation or affect task completion.

To support these requirements, Zhao described an infrastructure direction based on a global real-time transport network, distributed cloud-edge-device joint inference, real-time orchestration and full-duplex interaction, allowing models and agents to coordinate in real-world settings. Shengwang’s conversational AI engine uses full-duplex interaction as a core element to connect models, agent runtimes and users. The company is also improving developer tools and kits, including SDKs, Skills, CLI tools and Recipes, to help developers build agents with complete logic and execution capability, according to Leiphone.

Zhao said conversational AI is entering customer service, education, AI companionship and smart wearables. Whether AI creates value depends on whether a model can talk, understand needs, communicate naturally and complete tasks, he said. Nvidia chief architect and AOMedia Video Codec Working Group co-chair Andrey Norkin also spoke at the event about AV2 video codec technology.

In the first roundtable, moderated by Shengwang co-founder Tony Wang, Eugene Wei, Pan Luan and Hua Jianxin discussed whether personal agents could become the next personal entry point. Pan Luan said AI product value is shifting from capturing user attention to helping users complete tasks more efficiently, and future personal agents should take on clear, time-consuming tasks rather than only respond to explicit commands. Hua Jianxin said the next entry point may not belong to one device; glasses, earphones and other hardware that can sense the environment and execute tasks could all have opportunities. Eugene Wei focused on context: whether an agent can understand a user’s interests, communication records and social relationships will affect its value. He suggested watching daily user-agent interactions and whether users approach agents more often than agents approach users.

In the second roundtable, moderated by Shengwang chief commercial officer Cui Bo, Chen Wei, Geng Xianglei, Cai Guoxiang and Wu Kan discussed conversational AI from embodied intelligence, smart wearables, enterprise marketing and AI recruiting. Chen Wei said embodied intelligence must turn human intent into robot action, with robots combining visual, voice and status information to perceive, plan and decide in physical environments and adjust as conditions change. Cai Guoxiang said natural conversation is a core competitive requirement for AI glasses; low latency, stable real-time interaction and understanding of semantics and long-term context can make users feel less like they are talking to AI. He cited Rokid’s use of Shengwang RTC link services. Geng Xianglei said AI voice agents in enterprise settings must address latency, interruption timing and human-likeness, and must combine industry knowledge and human collaboration for business delivery. Wu Kan said conversational AI can help recruiting reach candidates at scale and improve service efficiency rather than simply replace people, but business closed loops require industry data, long-term agent memory and evaluation around interviews and onboarding.