🧩 LimiX-2: Towards General Structured-Data Intelligence

One pretrained model. Multiple structured-data tasks. #1 across three major tabular benchmarks.
LimiX-2 is a pretrained foundation model for structured data that can support
🎯 Classification
📈 Regression
🧩 Missing-value imputation
🔗 Causal skeleton recovery
— with no task-specific parameter updates for downstream prediction tasks.

🏆 #1 across all three benchmarks

LimiX-2 takes the top spot on all three broad tabular benchmarks we evaluate:
🏆 TabArena — 1935 Elo
🏆 TALENT — 1506 Elo
🏆 BCCO — 1432 Elo
LimiX-2 ranks #1 overall across all three, ahead of leading tabular foundation models and strong dataset-specific systems.
And the gains continue as we scale the model from 12.5M → 406.2M parameters: performance improves consistently across all five evaluated scaling series, with no clear sign of saturation within the measured range.
So how can one pretrained model handle such different structured-data problems?

A table is more than a target column.

Most tabular foundation models are organized around one question:
given ( x ), predict ( y ).
LimiX-2 takes a broader view:
Can a model learn the relationships among variables in a table, so that different unknown quantities can be inferred from different observed ones?
We introduce Contextual Mechanism Networks (CMNs), a new paradigm that shifts structured-data in-context learning from target-centric prediction toward context-dependent joint modeling.
Instead of focusing only on
$$p(y\mid x,D_{\mathrm{c}\mathrm{o}\mathrm{n}\mathrm{t}\mathrm{e}\mathrm{x}\mathrm{t}})p(y\; \backslash mid\; x,\; D\_\{\backslash mathrm\{context\}\})$$
CMNs aim to capture the broader dependency structure behind
$$p(x,y\mid D_{\mathrm{c}\mathrm{o}\mathrm{n}\mathrm{t}\mathrm{e}\mathrm{x}\mathrm{t}})p(x,y\; \backslash mid\; D\_\{\backslash mathrm\{context\}\})$$
LimiX-2 instantiates this idea with Context-Conditional Masked Modeling (CCMM).
During pretraining, different variables are masked under different observation patterns, turning the same table into many conditional prediction problems. Rather than learning only how to predict one designated target, the model learns to infer different variables from different available evidence.
LimiX-2 is pretrained exclusively on synthetic datasets generated from diverse structural causal models (SCMs), spanning different graph structures, functional mechanisms, distributions, and observation processes.
This broader training objective gives a single pretrained model a unified basis for classification, regression, and missing-value reconstruction.
Beyond predictive performance, we also find that LimiX-2's learned feature attention contains structural information that can be used for causal skeleton recovery—suggesting that learning conditional relationships across variables can capture structure beyond a single supervised target.
This points to a broader possibility:
structured-data foundation models may move beyond predicting predefined targets toward learning reusable mechanisms of how variables relate and interact.
** A table is more than a target column. Prediction may be only the beginning.**
The next question is:
How far can general structured-data intelligence scale?
LimiX-2 is just the beginning. Follow the project as we push toward larger models, broader capabilities, and more general structured-data intelligence. 🚀
📄 Paper — https://arxiv.org/abs/2609.17488
💻 Code — https://github.com/limix-ldm-ai/LimiX
🤗 Model — https://huggingface.co/stable-ai/LimiX-2
Reproductions, independent evaluations, integrations, and feedback are very welcome.

Great work on general structured‑data intelligence! Shifting from target‑centric prediction to joint mechanism‑oriented modeling looks like a promising direction for tabular foundation models. Impressive results across TabArena, TALENT and BCCO benchmarks. Really appreciate that the model also shows causal skeleton recovery capability beyond pure prediction performance. Looking forward to seeing further real‑world validations!