A two-year-old startup just got published in Nature by teaching AI to ask questions, not scale up

On May 28, 2026, Nature Communications published a paper on AI-based risk prediction for kidney cancer patients. The paper's co-first author is Wang Yatian, co-supervised by Wiener Intelligence CEO Liu Qifeng and HKUST's Luo Wenhan. But the paper is not just a medical AI milestone. It is a validation of a controversial thesis: that the most valuable training data is not harvested from the internet but generated adversarially by the model itself. Anthropic Launches Claude Science, an AI Workbench for…

Wiener Intelligence becomes the first Chinese data generation company, and only the fourth AI startup globally, to appear in a Nature journal, following DeepSeek and ModelBest among Chinese firms. The company was founded less than two years ago. DeepSeek's DSpark just fixed the two things that held…

The clinical problem

The research addresses a real surgical dilemma. Partial nephrectomy preserves kidney function but carries higher surgical risk. Radical nephrectomy is safer surgically but sacrifices the entire kidney. Surgeons lack reliable tools to predict which patients will suffer rapid renal function decline after surgery, the very complication that makes the decision agonizing.

The team built the Rapid GFR Decline Prediction Model (RDPM), which reframes the problem from short-term point estimation of postoperative eGFR to long-term functional risk stratification. The model uses multimodal multi-head cross-attention mechanisms to fuse 3D imaging data with clinical variables. Critically, the contralateral kidney cortex and medulla are automatically segmented by a UNest model, then reviewed by a physician, a hybrid approach that acknowledges the limits of fully automated segmentation in clinical settings. Microsoft's new platform gives scientists a governed…

Trained and validated on a multicenter cohort of 1,621 patients across 15 institutions, the RDPM achieved external test AUC of 0.788 to 0.873. Those numbers are competitive, though not revolutionary. The real innovation is in the how.

Reasoning data: the paradigm behind the numbers

Wiener Intelligence's core claim is that the AI industry has been optimizing the wrong variable. Most labs chase larger models and more parameters, assuming that scale automatically yields causality. The Hong Kong startup argues instead for reasoning data generation, a paradigm in which the model generates both questions and answers alongside their chain-of-thought reasoning, creating challenge sets that force adversarial, causality-rich knowledge organization. How Local LLMs Like Gemma and Qwen Are Taming Open…

The format is cQrA, context, Question, reasoning, Answer, a quadruple designed to teach models not just to answer questions but to ask them. Liu Qifeng, also a visiting professor at HKUST, told the team: The goal is to train AI models to become agents capable of autonomous learning. Not just answering machines.

This approach has already been deployed in value alignment and safety, financial insurance, Hong Kong government services, and competitive sports, domains where manual annotation is expensive and parameter scaling offers diminishing returns. The company claims industrial-grade precision without massive manual annotation or parameter scaling.

Why this matters beyond the OR

The Nature publication's significance extends beyond oncology. If Wiener Intelligence's reasoning data paradigm generalizes, it could reshape how AI systems are trained for high-stakes, data-scarce domains. The current industry consensus, embodied by DeepSeek, GPT, and Gemini, holds that scale is the primary driver of reasoning capability. Wiener Intelligence's counterpoint: the structure and provenance of the training data matters more than the parameter count. Ai2's EMO makes modular AI emerge from data, not human rules

The paper provides concrete evidence that a small company with a narrow compute budget can produce publishable scientific results across highly heterogeneous domains. That is a direct challenge to the scaling orthodoxy that dominates AI labs today.

The skepticism question

Skeptics will note that a 0.788 to 0.873 AUC, while respectable, does not represent a step change in predictive performance. The paper's real contribution is the data generation method, not the medical outcome. Whether the cQrA approach can be validated independently by other labs, and whether it scales to problems beyond the relatively narrow surgical risk domain, remains open. Ifbench: the new benchmark testing AI instruction following

But the fact that Nature Communications accepted the paper suggests the peer reviewers found the methodological contribution substantive. For a two-year-old startup from Hong Kong to clear that bar is itself a signal.

The next test will be whether Wiener Intelligence can replicate this result in another domain, perhaps one of the other high-stakes fields where it claims deployment. If it does, the conversation about what drives AI capability will have a new, very practical data point.