Open LimiX Studio

Stable AI / Research

Advancing structured-data intelligence

We study how models learn reusable relationships within structured data across tasks, datasets, and observation patterns.

SCALING × CONDITIONAL MODELING × STRUCTURE

Thesis

A table is a system of conditional questions

p(y | x, Dcontext)p(x, y | Dcontext)

Tables carry structure across both variables and observations. Prediction is one conditional question among many.

General models must learn relationships that can be reused across targets, datasets, and observation patterns, rather than fitting only one fixed target.

Research questions

Four questions guide our research

01 — Scaling

How does performance scale?

In the LimiX-2 scaling study, six model sizes from 12.5M to 406.2M parameters show positive log-linear trends across five reported evaluation series.

  • 12.5M → 406.2M
  • No reported saturation within the evaluated range

Report-scoped research finding

Observed LimiX-2 Log-linear fit
176612.5M178525M183850M1851100M1904200M1935406.2MMODEL PARAMETERS · LOG SCALE
Reported TabArena Elo across model scales
02 — Conditional modeling

Can one model answer many questions about the same table?

Different masks over the same table connect prediction, reconstruction, and broader conditional inference.

Same table
  1. 01
    Target column maskedPredict outcomes
  2. 02
    Missing cells maskedRecover values
  3. 03
    Arbitrary cells maskedAnswer conditional queries
The same table supports three separate conditional queries. Each mask is applied independently.
03 — Synthetic pretraining

How can synthetic mechanisms support generalization?

LimiX-2 is pretrained on generated tables spanning graph structures, functional mechanisms, and observation processes.

  1. 01

    Hyperparameter sampling

    Nodes · features · samples · task

  2. 02

    DAG generation

    Graph structures and motifs

  3. 03

    SCM propagation

    Functions · noise · aggregation

  4. 04

    Feature & target sampling

    Observed features · target

  5. 05

    Task adaptation

    Classification · regression

Five-stage synthetic data generation process for pretraining.
04 — Causal awareness

Does feature attention encode causal structure?

The report probes LimiX-2’s causal awareness by treating each variable in turn as the prediction target. Feature-to-target attention scores are thresholded to recover an undirected causal skeleton, which is evaluated against ground truth using F1 and SHD.

Attention probe · research finding

0.7972Mean F1 across six causal-discovery datasets
6 / 6Highest F1 among compared methods
5 / 6Lowest SHD among compared methods
Report-scoped causal-awareness probe. Reported F1 range: 0.7013–0.9385. The recovered structure is an undirected causal skeleton, not a directed causal graph or a released prediction interface.

Selected publications

Research behind LimiX

Technical reports, peer-reviewed work, code, and model resources from the LimiX research trajectory.

01
2026 · Technical report

LimiX-2: A Contextual Mechanism Network Towards General Structured-Data Intelligence

Introduces Contextual Mechanism Networks and a joint perspective on prediction, imputation, and structural relationships.

Technical report
02
2026 · Research paper

LimiX-2M: Mitigating Low-Rank Collapse and Attention Bottlenecks in Tabular Foundation Models

Studies efficient tabular foundation models in a compact two-million-parameter architecture.

ICML 2026
03
2025 · Technical report

LimiX: Unleashing Structured-Data Modeling Capability for Generalist Intelligence

Presents the first LimiX model and a generalist approach to structured-data modeling.

Technical report

Research collaboration

Work with us on the next questions

Discuss a research collaboration

Open research resources