From AI Build AI to True RSI: Inside the Zhangjiang Summit on AI Self-Evolution

Frontier Insights: Deconstructing AI Self-Evolution Practice Paths

The "Triple T" Zhangjiang Top Meeting series technical salon was held on September 23 at the Zhangjiang AI Innovation Town, jointly organized by Machine Heart and Zhangjiang AI Innovation Town Eco Company. Themed "AI²: When AI Learns to Self-Transcend," the event focused on frontier research and practical progress in self-evolving AI through six thematic presentations and a roundtable discussion, exploring technical paths, key bottlenecks, and breakthrough conditions across R&D automation, code models, agent systems, embodied intelligence, and scientific discovery.

Chen Siheng (Shanghai Jiao Tong University): Co-evolution of Agent-Environment-Data-Model Driven by Frontier Research

Chen Siheng, Associate Professor at the School of Artificial Intelligence, Shanghai Jiao Tong University, presented "Frontier Research Driving Recursive Self-Evolution of Agent-Environment-Data-Model." He argued that the core of recursive self-evolution is enabling AI to gradually reduce dependence on humans and achieve autonomous iteration and capability leaps. His proposed "Agent-Environment-Data-Model" co-evolution roadmap currently focuses on using AI to autonomously conduct frontier scientific research, continuously generating, screening, and precipitating high-quality data, which then feeds back into model capabilities to form a closed loop. Frontier research features high difficulty, verifiability, and scalability, enabling continuous traction of intelligence toward higher levels.

Yang Jian (Beihang University): Self-Evolving Code Large Models

Yang Jian, Associate Professor and Outstanding Youth Scholar at Beihang University, shared "Self-Evolving Code Large Models." He stated that code large models are a more grounded scenario for RSI. The core lies not only in optimizing the Agent Harness but in achieving Model RSI — letting the model continuously improve itself through execution, verification, and correction. Code tasks possess closed-loop environments and objective verification mechanisms, making them suitable for long-horizon Agentic RL. Future evolution can proceed through three layers — tasks, Harness, and Model — gradually connecting data, environment, and model into a full-element closed loop, extending from post-training to evaluation and deployment, pushing AI R&D processes toward automation and self-evolution.

Yang Ling (Peking University / PhAI Labs): From RSI to Discovery Intelligence

Yang Ling, Assistant Professor at Peking University, Postdoctoral Researcher at Princeton AI Lab, and Co-founder of PhAI Labs, presented "From RSI to Discovery Intelligence." He traced the research evolution from code RSI and agent interaction learning to scientific discovery intelligence. He argued that scientific research features long horizons, executability, verifiability, and access to real environmental feedback, making it a key scenario for exploring continuous learning and recursive self-improvement in AI. Future AI will continuously accept real evidence tests through autonomously discovering high-value problems, proposing hypotheses, designing and executing experiments, and converting research experience into transferable capabilities, pushing the upper bound of general intelligence via more complex and open scientific scenarios.

Zhou Xuanhe (Shanghai Jiao Tong University / Shanghai AI Lab / Theseus Labs): Judging Genuine RSI

Zhou Xuanhe, Tenure-track Assistant Professor at Shanghai Jiao Tong University, Dual-appointed Assistant Researcher at Shanghai AI Lab, and Youth Scholar at Beijing Academy of AI, referenced the Theseus Labs report "The Last AI Built by Humans: Toward Genuine Recursive Self-Improvement." He emphasized that true RSI should not be limited to local optimization of models on fixed tasks but must form a higher-level "outer loop" where AI autonomously discovers problems, proposes solutions, builds feedback, and verifies results, closing the loop among tasks, data, environment, and model. Based on this framework, Theseus Labs continues exploring environment evolution loops, data self-reasoning, and expanding model frontier boundaries, with new research conclusions and technical reports forthcoming.

Shi Ye (ShanghaiTech University / YesAI Lab / Shunshi Technology): Embodied Intelligence Self-Evolution

Shi Ye, Researcher, Assistant Professor, and PhD Supervisor at ShanghaiTech University, Head of YesAI Lab, and Founder of Shunshi Technology, presented "Embodied Intelligence Self-Evolution: World Models, Policy Learning, and Executable Experience." He highlighted unique challenges: data scarcity, complex real environments, and difficulty obtaining feedback. The core is not pursuing general generalization like large models but enabling robots to quickly adapt to new environments and tasks while continuously learning. Around "what to learn, where to learn, how to update, how to deploy," he proposed building interactive, learnable environments via world models, combining simulation, real interaction experience, and reinforcement learning to form a closed loop from experience generation, policy update, to real deployment, driving continuous evolution of embodied intelligence.

Zhang Haotian (Valhalla Technology): AI for AI4Science — Self-Evolving AI4Science Models

Zhang Haotian, Founder and CEO of Valhalla Technology, shared "AI for AI4Science: Self-Evolving AI4Science Models." He described the shift from "AI-assisted research" to "AI training AI." The core is enabling AI to autonomously complete scientific task diagnosis, hypothesis generation, experimental verification, and evidence precipitation. In small-sample, high-heterogeneity scenarios like biopharma, AI must not only automate data processing but also validate hypotheses through single-axis experiments, forming reusable experiential memory that continuously feeds back into model and task iteration, ultimately achieving self-evolution of scientific intelligence.

Viewpoint Clash: How Far Is True RSI?

The roundtable "AI Build AI: How Far Is True Self-Evolution?" was moderated by Zhu Ge Mingchen (Recursive founding member, KAUST PhD) with panelists Yang Jian, Yang Ling, Zhou Xuanhe, Shi Ye, and Zhang Haotian. Discussions centered on the definition and criteria of self-evolution, how models, data, agents, and environments form synergistic closed loops, and challenges in digital worlds, embodied intelligence, and AI for Science — including data scarcity, feedback acquisition, long-horizon tasks, experimental verification, and experience reuse. They also explored the judgment, environmental adaptation, and model self-improvement capabilities needed for genuine autonomous iteration, envisioning paths from local optimization to higher-order autonomous evolution.

Yang Jian: AI Coding is the most likely scenario to achieve RSI first, but true model self-evolution requires connecting model, data, infra, and tool calling beyond Harness-level optimization. The next key is stronger long-horizon task, problem discovery, and autonomous solving capabilities, gradually reducing human involvement.

Yang Ling: True self-evolution should not just solve given problems better but possess problem discovery, definition, and innovation capabilities. In AI for Science, mismatched wet/dry experiment speeds, sparse feedback, and data/environment construction must be solved to form a complete research self-evolution loop.

Zhou Xuanhe: RSI's core is moving AI from completing fixed tasks to autonomous exploration and self-improvement, ultimately forming its own judgment and "taste."

Shi Ye: Embodied RSI hinges on converting real-world failure experiences into reusable capabilities. Rather than relying solely on large models, combining VLA, Action Models, small models, and Harness through limited real interactions to continuously precipitate experience accelerates robot learning and decision-making.

Zhang Haotian: AI for Science breakthroughs require moving from "AI-assisted research" to "AI training AI." Multimodal models can alleviate scientific data sparsity, but true RSI depends on whether AI can form judgment surpassing expert experience and autonomously optimize models and research workflows.

Conclusion

Chen Heng, Vice President of Zhangjiang Group and Professor-level Senior Engineer, summarized: The Zhangjiang "Top Meeting" aims to gather young scientists and industrial innovation forces for frontier idea and technology collision. The Zhangjiang AI Innovation Town will continue building high-quality exchange platforms, turning the Zhangjiang Twin Towers from architectural landmarks into industrial and technological landmarks.

From "Can AI self-evolve?" to "How to make self-improvement continuous?" technical exploration is moving toward more concrete mechanisms and verification. This salon brought together research and practice from different fields, fully presenting progress, divergences, and open problems in self-evolving AI, providing direction for the next phase of breakthroughs. Continued collaboration between research and industry around data, tools, experimental environments, and real needs is expected to extend onsite thinking into new research and practice, advancing self-evolving AI from local exploration to more solid capability accumulation.