Youth Voices Conclude WAIC: Pushing the Talent Ceiling and Shaping AI’s Next Phase

On the first day of the 2026 World Artificial Intelligence Conference, the "Pioneer Youth Talk" organized by Machine Heart and co‑hosted by Shanghai Zhangjiang AI Innovation Town concluded successfully. The event showcased a record‑high density of youth talent and gathered Best Paper authors, multi‑billion‑dollar founders, top scholars, and core technologists. It employed an immersive world‑café rotation format to deconstruct industry iteration questions and surface frontier technical insights.

The Shanghai Municipal Talent Bureau presented a deep dive into Shanghai’s high‑level sci‑tech talent policies, covering overseas talent settlement guarantees, post‑doctoral special funding, and key industry talent rewards, illustrating the city’s robust talent‑development ecosystem.

The Shanghai Economic and Information Commission then outlined the AI industry’s high‑quality development layout, promoting compute, data, R&D conversion platforms, specialized pilot carriers, and cutting‑edge application scenarios, thereby demonstrating a complete AI ecosystem and full‑chain empowerment system.

Following the policy briefings, Hong Kong University MMLab PhD Chen Tianxing (also CTO of Xspark AI) hosted the world‑café session. Forty youth elites rotated through tables discussing Self‑Improve Agent, world models, AI4S and other cutting‑edge, controversial topics, fostering cross‑disciplinary debate and generating original viewpoints.

Shanghai Jiao‑Tong University associate professor Yang Yi and Liu Weiwen presented on "Self‑Improve Agent: Theoretical Limits and Engineering Realization". They described the agent as comprising three co‑evolving modules—model, runtime framework, and data—requiring environment perception, self‑identification of capability gaps, data generation, task‑trajectory evaluation, and iterative optimization. Evaluation hinges on clear task definition and feedback. They emphasized knowledge‑boundary training, experience retention via model‑framework co‑evolution, and highlighted safety risks such as alignment drift, recommending verification steps during iteration rather than post‑hoc fixes.

Galaxy General Robot co‑founder and large‑model lead Zhang Zhizheng, together with Hong Kong University of Science and Technology co‑founder Chi Xiaowei, explained how world models empower embodied intelligence. They identified three pathways: generating training data, serving as reinforcement‑learning environments, and learning physical representations. Physical realism is essential. Comparing WAM and VLA, they noted VLA’s stronger interpretability but slower inference, while WAM’s training paradigm remains exploratory. They warned of unresolved issues such as video generation, reward‑function design, and limited coverage of complex micro‑physics.

Qwen researcher Qiu Zihan and Tsinghua University PhD Wang Shaowen examined "Boundaries of Large‑Model Capabilities". They pointed out hard bottlenecks in short‑term memory and long‑sequence processing, limiting rapid human‑like recall and long‑task deployment. They argued that industry‑level data‑feedback loops enable continual iteration, yet models still suffer from reward speculation and anti‑intuitive flaws, which can only be mitigated by scenario‑specific constraints. Future risks include excessive execution speed, parallelism, and rapid propagation of harmful behavior, necessitating proactive scenario‑based safeguards.

Beijing University Shenzhen Institute assistant professor Yu Weihao and USTC post‑doctoral researcher Li Xin discussed "Multimodal Semantic Understanding for the Physical World and Efficient Generation". They stressed that physical priors such as gravity and friction must be embedded in representations. The community is split between a dual‑tower architecture with agent interaction and a unified representation trained on massive data and compute. Improving physical accuracy requires fine‑grained annotation and embedding physical laws, while balancing efficiency, quality, and evaluation metrics remains an open challenge.

Nanyang Technological University post‑doctoral researcher Wang Kun and Shanghai Jiao‑Tong University PhD Ma Ziyang presented "Next‑Generation Multimodal Interaction: Context Awareness and Non‑Verbal Reasoning". They outlined three research directions: early fusion of native multimodal pre‑training, encoder‑free pixel‑waveform parsing, and streaming real‑time interactive systems. Applications include lifelong personal assistants and multimodal code generation, though safety frameworks for large‑scale native multimodal models are still lacking.

Oxford University post‑doctoral researcher Tang Shixiang and CEO Wang Xuanzhe explored "AI for Science: Foundational Paths and Paradigm Shifts". They argued synthetic data can only supplement, not replace, real experimental data due to theoretical approximation errors and data dependency. High‑throughput automated labs are needed to provide high‑quality data. Scientific agents could become independent verification platforms, but face bottlenecks in data scarcity and closed‑loop validation. Hallucination versus inspiration remains debated, and fine‑tuning large models works for literature review but falls short on numerical computation and causal reasoning, making cross‑disciplinary scientific AI still hard to realize.

In conclusion, youth constitute the core driving force behind AI innovation, poised to bridge academic breakthroughs with industrial deployment. Their sustained, original contributions will fuel high‑quality iteration of the global AI industry.