A three-month-old Beijing AI startup has raised nearly 400 million yuan by betting that the next competitive advantage in model development will be a system that improves its own research process. Chaoyan Intelligence, or Chaoyan Zhineng, announced angel and angel-plus rounds on September 16, according to 36Kr. The company is focused on recursive self-improvement, often abbreviated as RSI, a term for systems designed to use their own outputs to improve future research and development work.
The financing brings major Chinese investors into a research direction that sits between AI agents and automated science. Chaoyan says its system can analyze a problem, propose a method, write code, run experiments, and then use results from those experiments to continue iterating. These are company and media descriptions of its intended capability, not independently audited proof that the system can replace scientific teams.
The startup joins an already active debate about whether AI can move from answering questions to accelerating the work of AI research itself. EastFrontier examined that competition in June in its analysis of China’s race for recursive self-improvement. Chaoyan’s funding shows that Chinese investors are now backing an early-stage local company built directly around that premise rather than treating RSI as a feature inside a broader model platform.
Two Rounds Place Recursive Self-Improvement at the Core
36Kr reported that Chaoyan was founded in June 2026 and completed the two financings within roughly three months. The amount is described as nearly 400 million yuan, not a precise disclosed total. IDG Capital, LinkX Capital, and XtalPi co-led the angel round, while Zhongguancun Science City Fund, Shenzhen Capital Group, and Shanghai Future Industry Fund jointly led the angel-plus round.
Other investors cited in the angel round include Decent Capital, SEE Fund, Monad Ventures, and Winsoul Capital. The two rounds give the company a mix of venture investors, state-backed funds, and a technology company with a presence in computational drug research. The reports do not identify a valuation, and you should not infer one from the funding amount or investor list.
The company’s research framing is unusually specific. Rather than promising a general assistant for enterprises or consumers, Chaoyan is pursuing a self-evolving foundation-model system. RSI can mean many things in AI marketing, so the useful distinction is operational: the startup says it wants a model to participate in research loops by generating approaches, executing tasks, checking results, and reusing results in subsequent work. The difficulty lies in ensuring that a system does not merely produce plausible text about research, but generates methods and code that remain useful after testing.
A Young Tsinghua Researcher Leads the New Company
Chaoyan’s founder is Chen Yongchao, whom 36Kr and Beijing News describe as a 27-year-old assistant professor at Tsinghua University’s School of Artificial Intelligence and a former Google DeepMind researcher. Those biographical details explain why the deal attracted attention, but they do not validate the company’s technology. They do, however, place the startup within the growing overlap between Chinese university research, returning technical talent, and venture-backed model companies.
The reports also say Chaoyan and academic collaborators introduced ASI-Bench, intended to assess whether AI can conduct scientific research when humans do not supply the method. That is a more demanding target than conventional agent benchmarking. An agent can complete a narrow workflow if a person defines the objective and tool sequence. A scientific-research system needs to formulate and test approaches while recognizing that a passing code run is not necessarily a meaningful result.
The company’s approach has precedents in China’s broader interest in self-improving systems. EastFrontier previously reported on Recursive Superintelligence, founded by former Meta researcher Tian Yuandong. Chaoyan differs by arriving with a sizable early financing package and a research narrative centered on automated scientific iteration. The two companies nevertheless illustrate how “self-improvement” has become a distinct category for founders and investors, not just a capability claimed by general-purpose model makers.
The Investment Test Will Be Reproducible Research Value
The investment is not a verdict on RSI’s feasibility. Systems that autonomously write code and run experiments can compound errors if they optimize toward misleading test results, rely on incomplete information, or fail to distinguish a novelty from a useful discovery. For that reason, credible evidence will require reproducible outputs, transparent evaluations, and results that withstand expert review outside the company.
Chaoyan’s immediate advantage is that its research agenda is legible to investors: it is attempting to build tools that may shorten the cycle of model improvement itself. Its immediate challenge is equally clear. The company must convert a broad aspiration about self-evolving AI into measurable demonstrations that researchers can examine. The source reports do not establish a production deployment or a published benchmark result, so the financing should be seen as early backing for a research direction rather than proof of commercial maturity.
China’s model sector has often been described in terms of chips, training data, and parameter counts. Chaoyan’s raise directs attention to another layer: the workflow through which models are improved. If the company can show that automated loops produce useful and verifiable research results, it could influence how Chinese labs allocate talent and compute. If it cannot, the nearly 400 million yuan will still be a measure of how strongly investors believe the next AI race may be fought inside the research process itself.