LimiX-2 Tops Three Benchmarks in Structured Data Foundations

The structured-data foundation model LimiX-2, developed by a Tsinghua University-affiliated team, has taken first place on international tabular-data benchmarks. According to reporting by Meng Yao of QbitAI, the Chinese team has taken the lead in an area where Google, Amazon, and SAP have entered one after another. The three benchmarks are TabArena, TALENT, and BCCO. It is notable for securing first place in binary classification, multi-class classification, and regression.

LimiX makes the model understand the causal mechanisms behind the data

This is the development principle highlighted in the original article. The model is said to have 400 million parameters. It is attracting attention as a different path from large language models.

Background Behind the Rise of Structured-Data

Foundation Models

Structured-data foundation models are also called LDMs. They aim to directly learn the relationships, conditional dependencies, and generative mechanisms that exist among variables in tabular form. They are positioned as technologies for addressing core challenges on the production floor. They target something different from the semanticized knowledge handled by language models. They are characterized by directly confronting real-world measurements and business records.

In manufacturing, temperature, pressure, rotation speed, raw-material mix, equipment status, power consumption, and other factors change in parallel. The number of variables ranges from dozens to hundreds, and in some cases reaches thousands. In many cases, important laws are not documented. The focus of interest is not on individual values but on stable relationships among variables. There is a need to extract relationships that hold even when the environment changes.

There is also a method of converting tabular data into text and feeding it to language models. However, information compression occurs during the conversion process. In production management, minute differences can determine yield and safety judgments. Differences that are negligible in semantic space can have value in production space. This gap has heightened expectations for LDMs.

How It Took First Place on Three

International Benchmarks

It started with SAP’s early release of TabPFN-3.5 on September 15. Just hours later, on September 16, a team led by Tsinghua University PhDs announced their results. They reported that LimiX-2, with 400 million parameters, had taken first place on three major benchmarks. Reporting by Meng Yao of QbitAI states that it topped the official TabArena ranking. The tasks covered include binary classification, multi-class classification, and regression.

The developer is Wenjun Intelligence. Last year, the company released LimiX as the country’s first structured-data foundation model. LimiX is said to have long maintained a top position on the leaderboards. Afterward, it reportedly focused on accumulating foundational technologies rather than chasing rankings. LimiX-2 marks a return to first place as a result of those efforts.

The team has shown that it values returning to the top at key milestones rather than a one-time first place. It places reproducibility and sustainability of performance at the center of its evaluation criteria. It appears to aim to demonstrate the superiority of its design philosophy rather than one-off tuning. Its consecutive results on international benchmarks have reinforced that claim. Securing first place just after the tech giants’ entry also drew attention.

A Strategy Centered on Learning Causal Mechanisms

At the core of LimiX-2 is the modeling of relationships among variables. Its direction differs from row-wise modeling specialized in predicting a target variable. It puts front and center the learning of joint dependency structures across variables and across samples. It aims to extract the causal laws behind the data. Its goal is to capture relationships that are highly valuable yet difficult to discover.

A comparison with the representative PFN approach is presented. The approach represented by TabPFN emphasizes rapid adaptation when context and objectives are given. Its core goal is accurate prediction of Y for new tables. It has demonstrated high effectiveness in classification and regression. It has a track record of driving forward the development of tabular foundation models.

Meanwhile, questions on the production floor do not stop at the value of Y. The question is how a change in one variable propagates to other variables. There is a view that conventional target-variable-centered designs create capability limits. Many PFN-based systems compress data into high-level representations before using them to achieve objectives. LimiX-2 differs in that it learns the relational structure itself before compression.

Another distinctive feature is that it built its own high-quality automatic data synthesis infrastructure. It is designed to first expose the model to a sufficiently complex data world. It then follows a procedure of tackling real-world tasks afterward. This is seen as aiming to increase robustness to diverse generative mechanisms. It is a configuration conscious of raising the ceiling of causal modeling capability.

The Current Competitive Landscape and

Successive Entries by Tech Giants

Competition in 2026 is clearly accelerating. Google announced TabFM, incorporating tabular-data foundation models into its foundation-model lineup. Amazon launched Mitra. It is exploring transfer across tables and tasks with a single pre-trained model.

SAP’s moves are even more aggressive. It previously developed SAP-RPT-1. This year, it acquired Prior Labs, the company behind TabPFN. It reportedly invested more than 1 billion euros to expand its structured-data AI research team. A few days ago, it integrated TabPFN-3.5 Plus into SAP AI Core. It is said to be used directly for cash-flow forecasting, payment delays, supplier risks, and customer churn.

An area that was once a new direction in academia has become a foundational capability. Major global technology companies are all investing resources. The axis of competition has broadened from prediction accuracy to relationship understanding and transferability. Integration into companies’ core decision-making is advancing. The field has entered a development race premised on practical application.

The research is led by Professor Cui Peng of the Department of Computer Science at Tsinghua University. He serves as chief scientist of Wenjun Intelligence. He has long studied structured data and causal intelligence. He is regarded as one of the proponents of the LDM concept. He has explored extending scaling laws to structured data and improving causal modeling capability through CMNs.

A Complementary Role to Language Models

Large language models have expanded their capability boundaries over the past two years. They have encoded documents, code, papers, mathematical formulas, and annotated images and video. They are characterized by handling information that has gone through human understanding, selection, and abstraction. They form knowledge, reasoning, and generation capabilities from contextual relationships. From an information-theory perspective, they target reality that has already been semanticized.

Industrial systems contain data that remains outside the semantic layer. Measurements and state variables constitute a high-dimensional data space. They include laws that have not been organized in advance as rules or formulas. The stability of relationships matters more than identifying individual values. What is required is prediction of how the entire system will respond after a change.

For this reason, the two are positioned as parallel and complementary paths. While language modeling learns a represented world, LDMs enter high-dimensional space directly. They aim to understand laws from variable relationships themselves. The more complex the decision-making scenario, the greater the importance of LDMs. They have grown in presence as a capability component of the foundation-model ecosystem.

In practice, demands for accuracy, stability, and causality are high. Decision-making in industrial production and scientific research is a typical example. The value of LDMs becomes clear under these conditions. They are evaluated as a destination for analytical challenges that language models cannot fully handle. Expectations are gathering for them as a path deep into production.

Editorial Opinion

On short-term impact. The results are seen as changing the evaluation criteria for tabular foundations over the next three to six months. Not only prediction accuracy but also reproducibility of relational structures and transferability will become subjects of comparison. To counter SAP’s integration into business operations, Google and Amazon are also seen as rushing verification on practical metrics. Taking first place at the 400-million scale is evaluated as encouraging efficiency-focused development.

On the long-term perspective. In one to three years, designs centered on causal understanding are seen as spreading to production management and financial screening. The pattern where the quality of synthetic-data infrastructure determines performance will strengthen. Combined systems that pair language models with structured models are evaluated as becoming the standard. Rather than dominance by a specific company, the challenge is seen as becoming the framework for interoperability.

Questions from the editorial team. It is asked under what conditions the advantage of relational learning is reproduced. Whether topping public benchmarks leads directly to stability in real-world settings remains unverified. Is a third-party verification framework needed to detect bias in synthetic data. How to balance performance competition with operational responsibility is evaluated as the issue.

References

• “AGI新战场谷歌亚马逊巨头激战，杀出个中国LimiX-2赢了又赢”, by 梦瑶 — 量子位, 2026-09-18T12:22:47.000Z (ARR)
• Source URL: https://www.qbitai.com/2026/09/492175.html