Improved tabular models learn better from real-world data.
problem Tabular models perform poorly on real-world datasets when trained only on synthetic data.
method Continued pre-training on a curated set of real-world datasets.
result Real-TabPFN achieves superior predictive accuracy on 29 datasets.
Framework fine-tunes foundation models with semi-supervised learning for downstream tasks and latent spaces.
problem Training foundation models with limited labelled data.
method Mutual information decomposition for downstream and latent spaces, semi-supervised fine-tuning.
result Significant improvements in classification tasks under low-labelled conditions.
Tabular foundation models outperform other methods in conditional density estimation across various datasets.
problem Estimating the full conditional distribution of a response given tabular covariates, especially in settings with heteroscedasticity, multimodality, or asymmetric uncertainty.
method Benchmarked three tabular foundation model variants (TabPFN and TabICL) against six CDE baselines on 39 real-world datasets.
result Tabular foundation models achieve the best CDE loss, log-likelihood, and CRPS across all sample sizes, outperforming other methods.
Foundation models leak sensitive data in synthetic tabular data generation, especially LLaMA 3.3 70B.
problem Privacy leakage in synthetic tabular data generation using foundation models.
method Benchmarked three foundation models (GPT-4o-mini, LLaMA 3.3 70B, TabPFN v2) against four baselines on 35 real-world tables.
result Foundation models, especially LLaMA 3.3 70B, have the highest privacy risk in synthetic tabular data generation.
Framework tackles class imbalance and noisy labels in active learning.
problem Class imbalance and noisy labels in real-world datasets.
method Uses foundation model priors to select informative samples for active learning.
result Substantial annotation savings (over 50%) with preserved performance and robustness.
Foundation models improve on econometric benchmarks for forecasting volatility, but vary widely across models.
problem Comparing pretrained time series foundation models to econometric benchmarks for volatility forecasting.
method Systematic comparison of nine zero-shot TSFMs against eight econometric specifications on 50 assets across 3 markets and 3 horizons.
result Tiny Time Mixers (TTM) is the only model that consistently beats the Log-HAR benchmark, but performance varies widely across models.
Unified pipeline predicts equipment anomalies with high precision and reduced false positives.
problem Predicting equipment anomalies before failures using small models.
method Triplet Feature Fusion combining statistical, time-series, and text embeddings.
result Achieves high precision and reduced false positives.
GOTabPFN improves tabular model performance with compact tokenization for HDLSS data.
problem Making tabular models effective for high-dimensional, low-sample size data without retraining.
method Introducing Graph-guided Ordering with Local Refinement (GO-LR) and Neuro-Inspired Subunit Compression (NSC) to create compact meta-features.
result GOTabPFN improves stability and accuracy in tabular benchmarks with compact tokenization.
TabPFN-2.5 boosts tabular AI performance, especially for large datasets.
problem Improving tabular AI models for large datasets and complex feature sets.
method Advanced tabular foundation model with data augmentation and distillation techniques.
result TabPFN-2.5 outperforms tuned tree-based models and AutoGluon 1.4 on benchmark datasets.
Generative model creates synthetic unlabeled data for SSL.
problem Training SSL models without real unlabeled datasets.
method Meta-optimized synthetic samples generated from generative models.
result Synthetic samples improve SSL performance more efficiently than real unlabeled data.
Tabular in-context learners perform well on biomolecular tasks, but performance depends on the representation used.
problem Predicting biomolecular properties from limited labeled data.
method Evaluating tabular in-context learners on protein fitness regression and small-molecule classification tasks.
result Tabular in-context learners are competitive for protein fitness regression but not for small-molecule classification.
PFNs pre-train models on simulated data to predict class probabilities.
problem Training machine learning models on large datasets.
method Pre-train a fixed model on small simulated datasets and use it to infer class probabilities in-context.
result PFNs achieve state-of-the-art performance and improve with larger inference data.
Unified framework for optimal liquidation with small market impact and semimartingale strategies.
problem Optimal liquidation under small market impact and portfolio liquidation.
method Semimartingale strategies and convergence results for BSDEs with singular terminal conditions.
result Unified framework for embedding two common liquidation models and microscopic foundation for semimartingale strategies.
Transformer model outperforms classical methods in childhood anemia prediction across diverse countries.
problem Generalizing childhood anemia prediction models across different countries and data scarcity.
method Transformer-based tabular foundation model compared to classical supervised methods using DHS data.
result Transformer model achieves lower Brier score and ECE in low-data settings, outperforming classical models.
We analyze the Hessian spectra of large models up to 100B parameters.
problem Accurate Hessian spectra of large foundation models are difficult to obtain.
method We use shard-local finite-difference Hessian vector products and stochastic Lanczos quadrature.
result We produce the first large-scale spectral density estimates of foundation models.
Deep generative model discovers inhibitors for unknown targets.
problem Discovering novel inhibitor molecules for unknown drug targets.
method Deep generative framework trained on protein sequences, small molecules, and interactions.
result Micromolar-level inhibition observed for two out of four synthesized candidates, including activity against SARS-CoV-2 variants.
This paper establishes a theoretical foundation for super-models via domain adaptation.
problem Reducing computational and data costs in AI for small and medium-sized enterprises.
method Two-stage diffusion process modeling, including pre-training and fine-tuning stages, using the Uhlenbeck-Ornstein process.
result The generalization error of the fine-tuning stage is dominant in domain adaptation.
Survey of financial foundation models for diverse applications.
problem Challenges in applying general-purpose FMs to financial tasks.
method Review of financial foundation models (FFMs) in three modalities.
result Emergence of FFMs designed specifically for finance.
Revises Bayesian model averaging for foundation models.
problem Ensemble pre-trained and lightly-finetuned foundation models for improved classification performance.
method Introduces trainable linear classifiers and computationally cheaper model averaging scheme (OMA).
result Ensembled models can better predict on various datasets.
TabCF simplifies causal inference for distributional effects.
problem Estimating causal effects with unmeasured confounding.
method Control function regression using tabular foundation models.
result Accurate, fast, and transparent causal estimation of distributional quantities.
We develop the mathematical foundations of the stochastic modified equations (SME) framework for analyzing the dynamics of stochastic gradient algorithms, where the latter is approximated by a class of stochastic differential equations with small noise parameters. We prove that this approximation can be understood math…
Time series foundation models are well-calibrated, improving over baseline models.
problem Calibration of time series foundation models for practical applications.
method Systematic evaluations of five time series foundation models and two baselines, assessing calibration, prediction heads, and long-term forecasting.
result Time series foundation models are consistently better calibrated than baseline models and do not show over- or under-confidence.
AutoWS-Bench-101 evaluates automated weak supervision methods for diverse domains.
problem Limited applicability of weak supervision due to difficulty in designing labeling functions.
method Automates labeling function design using a small set of ground truth labels.
result AutoWS methods often require foundation models to outperform simple few-shot baselines.
Introduces foundation priors for using model-generated data in empirical research.
problem Using model-generated data as real observations in empirical research.
method Introduces foundation priors as an exponential-tilted, generalized Bayesian update of the user's primitive prior.
result Synthetic data reflects both model patterns and user's priors, enabling principled use in empirical work.
Combines foundation models with weak supervision to improve NLP and video tasks.
problem Leveraging weak supervision with foundation models without labeled data.
method Liger, a combination of foundation model embeddings and weak supervision techniques.
result Liger outperforms existing weak supervision methods by 14.1 points on benchmark NLP and video tasks.
Foundation models trained on collider data improve jet generation tasks.
problem Improving foundation models for jet generation tasks.
method Pre-training OmniJet-α model on AspenOpenJets dataset. result Pre-trained model improves performance on jet generation tasks with domain shift.
New measure assesses time series pre-training data quality without labels.
problem Challenges in collecting diverse pre-training datasets for time series classification.
method Contrastive-learning-based foundation model and contrastive accuracy measure.
result Contrastive accuracy correlates with model performance on downstream tasks.
Paper explores foundation models for dynamical systems using synthetic data.
problem Lack of synthetic data for dynamical systems training.
method Pretrained transformer model on synthetic dynamics functions sampled from RKHS.
result Pretrained model generalizes across various dynamical systems in simulations and hardware.
TempBalance boosts model performance with low data.
problem Low-data training and fine-tuning in model alignment.
method Inspired by HT-SR theory, TempBalance balances training quality across layers.
result TempBalance improves model performance as data decreases.
GAMformer bridges tabular models and interpretability, offering a single-pass approach.
problem Lack of interpretability in tabular foundation models like TabPFN.
method In-context learning for GAM shape functions, training on synthetic data.
result GAMformer performs comparably to other leading GAMs across various classification benchmarks.
Paper speeds up large foundation models for time series data.
problem Resource-intensive foundation models limit accessibility.
method Dimensionality reduction techniques, including PCA and neural network adapters.
result Up to 10x speedup and 4.5x more datasets fit on a single GPU.
We propose a practical and scalable Gaussian process model for large-scale nonlinear probabilistic regression. Our mixture-of-experts model is conceptually simple and hierarchically recombines computations for an overall approximation of a full Gaussian process. Closed-form and distributed computations allow for effici…
New method reduces fine-tuning cost for reused models.
problem Repeating fine-tuning costs with outdated foundation models.
method Portable Reward Tuning (PRT) trains a reward model to maximize the same loss function as fine-tuning.
result PRT achieves comparable accuracy to inference-time tuning with less inference cost.
This work aims to create a large-scale model for critical care time series data.
problem Lack of large-scale datasets and distribution shifts in critical care time series data.
method Harmonized dataset creation and transfer learning research.
result Established a foundation for large-scale multi-variate time series models in critical care.
Foundation models outperform supervised methods in time series forecasting across various operational regimes.
problem Lack of domain-specific training and ongoing maintenance in supervised learning for time series forecasting.
method Evaluation of foundation models against standard supervised approaches across four operational regimes: periodic, physically constrained, stochastic, and demand forecasting.
result Foundation models are optimal for cold-start or long-tail scenarios and perform well in domains with transferable periodic structures.
Foundation models improve wage gap decomposition by capturing omitted career history factors.
problem Estimating wage disparities using incomplete career history data.
method Fine-tuning foundation models to mitigate omitted variable bias and estimate wage gaps.
result Foundation models can decompose gender wage gaps more accurately than traditional econometric methods.
Paper proposes verifier engineering for improving foundation models.
problem Challenges in providing effective supervision signals for foundation models.
method Leverages automated verifiers to perform verification tasks and deliver feedback.
result Verifier engineering can enhance foundation models' capabilities.
Foundation models improve time series prediction reliability, especially with limited data.
problem Improving time series prediction reliability with limited data.
method Comparison of Time Series Foundation Models (TSFMs) with traditional methods in conformal prediction.
result TSFMs provide more reliable conformalized prediction intervals and more stable calibration with limited data.
Train a lightweight carry-on model on existing LLMs for faster customization.
problem Customizing large language models for specific tasks is computationally expensive.
method Train an additional branch of transformer blocks on the final-layer embedding of pretrained LLMs, then merge them with a carry-on module.
result Training a 100M carry-on layer requires less than 1GB GPU memory, making it scalable and affordable.
Foundation AI model outperforms traditional VaR methods in forecasting.
problem Forecasting Value-at-Risk (VaR) for financial returns.
method Time-series foundation AI model, pre-trained on diverse datasets, fine-tuned for specific quantiles.
result Fine-tuned foundation model consistently outperforms traditional methods in actual-over-expected ratios.
Pretrained time-series models outperform train-from-scratch baselines in financial return forecasting.
problem Financial return forecasting
method Pretrained time-series foundation models
result Pretrained TSFMs dominate the ranking distribution, accounting for 8 of 10 task-level wins.
Mantis improves time series classification using a transformer model trained on synthetic data.
problem Insufficient application of foundation models to time series classification.
method Pre-trained transformer model on synthetic data, enhanced test-time methodology.
result Mantis achieves state-of-the-art performance across diverse datasets.
Paper introduces TS-GPT for engineering time series forecasting.
problem Engineering time series require causal operations, unlike linguistic data.
method Innovations representation theory, Generative Pre-trained Transformer.
result TS-GPT effectively forecasts real-time locational marginal prices.
Study evaluates interpretability of time series foundation models' latent spaces.
problem Improving interpretability of latent spaces in time series models for visual analytics.
method Evaluated MOMENT family of transformer-based models on five datasets, fine-tuning for performance.
result Fine-tuning improved latent space clarity but limited interpretability remained.
Density expansions for hypoelliptic diffusions (X1,...,Xd) are revisited. In particular, we are interested in density expansions of the projection (XT1,...,XTl), at time T>0, with l≤d. Global conditions are found which replace the well-known "not-in-cutlocus" condition known from heat-kernel asymptot…
Study shows cliff-learning in transfer learning from foundation models.
problem Data-scaling of transfer learning from foundation models in low data regimes.
method Investigation of cliff-learning phenomenon through foundation-model analysis and toy models.
result Cliff-learning reflects compatibility between priors and tasks.
GIFT-Eval benchmarks time series forecasting models across diverse datasets.
problem Lack of comprehensive benchmarks for evaluating time series foundation models.
method Developed GIFT-Eval, a benchmark with 23 datasets, 177 million data points, and 144,000 time series.
result Promotes evaluation of foundation models across various domains and frequencies.
VMoER improves uncertainty quantification in MoE layers for scalable foundation models.
problem Uncertainty quantification in large-scale models like MoE layers.
method Structured Bayesian approach with amortized variational inference over routing logits and temperature parameter inference.
result Improves routing stability, reduces calibration error, and increases AUROC by 12%.