A Chinese team released LimiX-2, a structure data foundation model, achieving first place across TabArena, TALENT, and BCCO benchmarks.
According to 量子位, a Chinese team led by Tsinghua PhD researchers released LimiX-2, a structure data foundation model with 400M parameters, which achieved first place across TabArena, TALENT, and BCCO benchmarks in binary classification, multi-class, and regression tasks. The team introduced Contextual Mechanism Networks (CMNs) and a context conditioning mask modelling (CCMM) approach, enabling cell-level data representation and automated synthetic data generation, and secured top Elo scores on TabArena (1917) and TabArena regression (2206).
Key facts
- 01LimiX-2 is a 400 million-parameter structure data foundation model that achieved first place across TabArena, TALENT, and BCCO benchmarks in binary classification, multi-class, and regression tasks.
- 02The model introduces Contextual Mechanism Networks (CMNs) and context conditioning mask modelling (CCMM), prioritizing learning relationships among variables over single-task prediction.
- 03An automated high-quality data synthesis engine is used to expose the model to more diverse data structures and support cross-variable and cross-sample interactions at the cell level.
- 04On TabArena, LimiX-2 ranks first in all four metrics with an Elo of 1917; on TabArena regression, it also ranks first with an Elo of 2206.
- 05The development aims to move beyond predicting a single target to understanding stable variable relationships across environments and generating data-driven causal structures.
AI-generated from the sources below. Always check the originals.
How each country tells it
So far our sources show coverage from one country. Perspectives appear once we find reports from a second country.
Sources
Summaries are AI-generated from the linked sources and may contain errors; always check the originals. We summarise and link; we never republish articles. Photos come from openly licensed libraries, official publicity material and brand logos, credited to their sources. If you own an image and want it credited differently or removed, email info@coda.news and we will act promptly.