Research Is the Engine: 27-Year-Old Tsinghua Professor's RSI Startup Apex Intelligence Raises ~$50M in Two Months

Recursive self-improvement (RSI) has become the most crowded thesis in AI right now — and one of the youngest founders in the field just landed one of the largest angel rounds in China to chase it. On September 16, 2026, Beijing-based Apex Intelligence (超衍智能) announced it had closed nearly RMB 400 million (about $50 million) across its angel and angel+ rounds, completed within roughly two months of the company’s public launch. The founder is Chen Yongchao, a 27-year-old who this year became the youngest assistant professor in the history of Tsinghua University’s School of Artificial Intelligence.

Who is building it

Chen’s resume reads like a tour of the frontier’s research benches. He completed his doctorate through the joint Harvard–MIT training program, and held research stints at Google Research, Google DeepMind, Microsoft and IBM before returning to China. In January 2026, he says, he stood at a crossroads: a job offer from DeepMind and co-founder invitations from several overseas startups working on AI-driven scientific research on one side; on the other, a startup window he had been waiting two years to open. He chose to return, founding Apex Intelligence while joining Tsinghua’s faculty — a dual track that mirrors how China’s AI ecosystem increasingly blurs academia and company-building.

The founding team is research-heavy by design. Core members come from ByteDance, Kimi (Moonshot AI) and Zhipu AI, bringing frontline large-model R&D experience. The company positions itself not as another agent-framework vendor but as a foundation-model company whose product is self-evolution itself: a general model that can independently run the full loop of research — conceiving ideas, retrieving literature, connecting GPUs, building environments, running experiments, and writing up the results — and then feed every research trajectory back into itself as training data.

The rounds

The financing is notable less for its size than for who wrote the checks and how fast it came together. The angel round was co-led by IDG Capital, Starlink (星连) Capital and XtalPi (晶泰科技), with follow-ons from Duxun Capital, Infinity Fund, Chuxin Capital and Yunxiu Capital — a mix of US-dollar funds, industrial capital and market-driven investors. The angel+ round was co-led by state-affiliated funds from China’s three biggest tech hubs: Zhongguancun Science City Fund (Beijing), Shenzhen Venture Capital, and Shanghai Future Industry Fund.

That structure tells its own story. A two-month, two-round close at this scale signals both intense investor appetite for the RSI narrative and the strategic weight Chinese cities are placing on “next-generation industrial intelligence.” The capital is earmarked for three things: foundation-model and compute infrastructure, sustained R&D into recursive self-improvement, and team expansion.

The claim that turns heads: 34 papers, two above 99% of humans

The most provocative data point in Apex’s story came in May 2026, before the company even announced itself. Chen says the team let its self-evolving AI system independently produce 34 papers and submit them to ACL Rolling Review, the unified review platform for ACL-series NLP conferences. Eleven of them scored above 3.0 in preliminary review — a bar roughly at doctoral-student level — and two scored higher than 99% of human researchers in the same pool.

Chen’s framing is cautious but pointed: “Many ideas are quite eye-opening. Some papers are almost at the level of doctoral students, which is beyond our expectation, but also reasonable, because this moment will come sooner or later.” The company also claims its system has completed a full proof of a 30-year-old majorization conjecture in mathematics — the kind of result that, if independently verified, would move RSI from narrative to evidence. Internally, Apex says it has accumulated tens of thousands of high-quality research trajectories, spanning expert trajectories, synthetic trajectories, and trajectories the AI generated autonomously.

“Research is the engine of RSI”

Apex’s differentiation argument in the increasingly noisy RSI field is a layered one. Agent-layer startups build harnesses — scaffolding that lets an existing model do research tasks. Apex instead targets recursive self-improvement at the foundation-model layer, turning each completed research task into training data that makes the model itself better at the next, harder task. Inside the company this is summarized as “Research is the engine of RSI.”

Chen draws a sharp distinction between today’s models and what he is building: current large models are stability-oriented — conservative, optimized to avoid mistakes across a hundred questions. A self-evolving model should be innovation-oriented: asked a question it cannot answer, it should generate a hundred candidate approaches, tolerate ninety-nine failures, and retain the one that works as a reusable capability. His uncomfortable observation: the better a model scores on general benchmarks, the worse it often performs at genuine innovation — a trade-off he believes requires different training methods (including adversarial self-play between idea generation and idea evaluation) and different architectures, not just more scaling.

On the timeline question, Chen is an S-curve thinker. The first S-curve — the classic Scaling Law era — is approaching saturation at its upper end. RSI may be the next S-curve, or a methodology for discovering many of them. “The times will not wait for you,” he says. “This is the same situation as not making large models two years ago and not making embodied intelligence one year ago.”

Context: a crowded, mostly unproven field

The round lands amid an RSI funding wave that is as much narrative as substance. Andrej Karpathy, former Meta FAIR research director Tian Yuandong, and former Google Chief Scientist Jeff Dean have all recently joined or backed RSI-focused ventures. Tian’s Recursive Superintelligence, notably, announced $650 million raised at a $4.65 billion valuation in May 2026 before delivering public results — a benchmark for how hot the category has become. A 35-author academic roadmap published this month (“The Last AI Built by Humans”) argued that almost everything marketed as “self-improving” today sits at level two of a five-level autonomy ladder at best.

That is the backdrop against which Apex’s numbers should be read. Nearly RMB 400M is large for an angel round but modest against the category’s giants; the two-month close and the three-city state-fund consortium say more about strategic positioning than about technical proof. The company says its research-automation product recently entered small-scale internal testing, and that its in-house model — mid-trained and post-trained on open-source foundations, with large parameter scale — trades some general-purpose strength for significantly stronger complete-research capability.

Why it matters

Three things make this story worth watching beyond the funding headline. First, it is a test of whether self-evolution can be a foundation-layer product rather than an agent-layer feature — most RSI bets so far are scaffolds on top of existing frontier models. Second, the ACL Rolling Review numbers, whatever their ultimate scholarly fate, are one of the few quantified public claims in a field dominated by roadmaps. Third, the investor coalition — dollar funds, an AI-pharma unicorn, and state capital from three cities — sketches the political economy of China’s next AI wave better than any policy paper: the money now follows anyone credible who claims to compress the research loop itself.

The honest caveats are equally clear. Review scores on a rolling platform are not acceptance, and two statistical outliers do not establish a capability. A math conjecture proof awaits formal verification. And the innovation-versus-stability trade-off Chen describes is a hypothesis, not a settled result. But as the RSI race accelerates on both sides of the Pacific, Apex Intelligence has just secured the resources to make its version of the claim testable — and one of the field’s youngest founders now has the compute to try.