Research themes

In real-world decision systems — from healthcare diagnostics to financial risk assessment — knowing how confident a model's prediction is can be just as important as the prediction itself. Traditional ML methods often produce poorly calibrated confidence scores, especially when data is scarce or imbalanced. Reliable uncertainty estimation not only increases trust but also enables downstream tasks like active learning and risk-aware planning.

• Develop metrics and benchmarks

  Evaluating confidence calibration in TFMs, such as Expected Calibration Error and Brier scores across diverse structured datasets.

  // R.DIR. 1

• Explore hybrid approaches

  Bayesian/meta-learning approaches to provide both epistemic and aleatoric uncertainty estimates.

  // R.DIR. 2

• Investigate uncertainty behavior

  Under distributional shift and small sample regimes — crucial for trustworthy AI systems.

  // R.DIR. 3

AI systems that communicate not just predictions — but confidence — making them safer for high-stakes real-world deployment.

Powerful models can outperform traditional methods, but their decisions must be understandable to earn user trust, especially in regulated domains like healthcare, insurance, and public policy. Unlike models for text or images, structured data explanations require tailored techniques that respect column semantics, dependencies among features, and the inherent logic of tabular relationships.

• Design interpretability frameworks

  Work with TFMs' in-context learning mechanisms, rather than as an afterthought.

  // R.DIR. 1

• Investigate counterfactual explanations

  Feature attribution methods that highlight how and why a model altered a decision.

  // R.DIR. 2

• Connect interpretability research

  to domain knowledge, making insights actionable for subject-matter experts.

  // R.DIR. 3

By making powerful tabular models transparent, this work lays the foundation for responsible, human-aligned AI that users can trust and act upon.

Progress in AI relies on benchmarks — shared tasks that push the state of the art and make comparisons fair and reproducible. In tabular ML, existing datasets are often limited (too clean, too small, or unrepresentative), masking performance gaps in real scenarios. As foundational tabular models evolve, so must our benchmarks: they should reflect realistic challenges such as noisy entries, mixed feature types, missing values, imbalance, and temporal structure.

• Curate and maintain living benchmark suites

  That grow with the community and include real industrial data patterns.

  // R.DIR. 1

• Introduce holistic metrics

  Beyond accuracy — calibration, fairness, robustness to noise, and computational cost.

  // R.DIR. 2

• Facilitate open benchmark competitions

  And leaderboards to unify the research community's efforts.

  // R.DIR. 3

Better benchmarks accelerate innovation, enable fair comparisons, and clarify where the next breakthroughs are needed.

Early foundational models have shown remarkable predictive power on small and medium tabular datasets, and recent versions push to hundreds of thousands of rows and thousands of columns. However, many real enterprise datasets are orders of magnitude larger and more complex — from nationwide medical records to global customer databases. Building models that scale without sacrificing accuracy or interpretability is a defining research challenge in structured AI.

• Innovate efficient attention

  And memory mechanisms that allow models to handle billions of rows without performance degradation.

  // R.DIR. 1

• Create feature summarization

  And context selection strategies that retain essential signals without overwhelming computation.

  // R.DIR. 2

• Explore hybrid architectures

  That combine foundation model pre-training with scalable model families.

  // R.DIR. 3

Scaling TFMs unlocks their use across industries and applications that depend on massive structured datasets — from energy systems to astronomical survey analysis.

As models power more decision-making systems, we must ask: how do they behave under data corruption, malicious manipulation, or unexpected inputs? Recent research shows that even sophisticated TFMs can be vulnerable to targeted test-time perturbations, and that such vulnerabilities can transfer adversarially across models. Moreover, trustworthiness isn’t only about resisting attacks — it’s about predictable, safe responses when data deviates from training conditions.

• Develop robust training

  And evaluation frameworks that stress-test models against structured perturbations and distribution shifts.

  // R.DIR. 1

• Investigate in-context adversarial defenses

  Where models adaptively reinterpret context under uncertainty.

  // R.DIR. 2

• Formalize safety criteria

  Tailored to structured domains, integrating them into deployment pipelines.

  // R.DIR. 3

Robust, safety-aware tabular models ensure reliability in fields where mistakes aren’t just inconvenient — they can be costly or dangerous.