Prior-data fitted networks (PFNs) have recently been proposed as a promising way to train tabular foundation models. PFNs are transformers that are pre-trained on synthetic data generated from a prespecified prior distribution and that enable Bayesian inference through in-context learning. In this paper, we introduce CausalFM, a comprehensive framework for training PFN-based foundation models in various causal inference settings. First, we formalize the construction of Bayesian priors for causal inference based on structural causal models (SCMs) in a principled way and derive necessary criteria for the validity of such priors. Building on this, we propose a novel family of prior distributions using causality-inspired Bayesian neural networks that enable CausalFM to perform Bayesian causal inference in various settings, including for back-door, front-door, and instrumental variable adjustment. Finally, we instantiate CausalFM and explicitly train models to perform in-context learning in these settings. We show that CausalFM achieves competitive in-context learning performance even when compared to baselines that are specifically trained for the task at hand. In sum, our framework can be used as a general recipe to train foundation models for various causal inference settings. In contrast to the current state-of-the-art in causal inference, CausalFM offers a novel paradigm with the potential to fundamentally change how practitioners perform causal inference in medicine, economics, and other disciplines.

LLM leaderboards often rely on single stochastic runs, but how many repetitions are required for reliable conclusions remains unclear. We re-evaluate eight state-of-the-art models on the AI4Math Benchmark with three independent runs per setting. Using mixed-effects logistic regression, domain-level marginal means, rank-instability analysis, and run-to-run reliability, we assessed the value of additional repetitions. Our findings shows that Single-run leaderboards are brittle: 10/12 slices (83\%) invert at least one pairwise rank relative to the three-run majority, despite a zero sign-flip rate for pairwise significance and moderate overall interclass correlation. Averaging runs yields modest SE shrinkage ($\sim$5\% from one to three) but large ranking gains; two runs remove $\sim$83\% of single-run inversions. We provide cost-aware guidance for practitioners: treat evaluation as an experiment, report uncertainty, and use $\geq 2$ repetitions under stochastic decoding. These practices improve robustness while remaining feasible for small teams and help align model comparisons with real-world reliability.

Modern causal discovery methods face critical limitations in scalability, computational efficiency, and adaptability to mixed data types, as evidenced by benchmarks on node scalability (30, $\le 50$, $\ge 70$ nodes), computational energy demands, and continuous/non-continuous data handling. While traditional algorithms like PC, GES, and ICA-LiNGAM struggle with these challenges, exhibiting prohibitive energy costs for higher-order nodes and poor scalability beyond 70 nodes, we propose \textbf{GUIDE}, a framework that integrates Large Language Model (LLM)-generated adjacency matrices with observational data through a dual-encoder architecture. GUIDE uniquely optimizes computational efficiency, reducing runtime on average by $\approx 42%$ compared to RL-BIC and KCRL methods, while achieving an average $\approx 117%$ improvement in accuracy over both NOTEARS and GraN-DAG individually. During training, GUIDE's reinforcement learning agent dynamically balances reward maximization (accuracy) and penalty avoidance (DAG constraints), enabling robust performance across mixed data types and scalability to $\ge 70$ nodes -- a setting where baseline methods fail.

Post-hoc calibration methods are widely used to improve the reliability of probabilistic predictions from machine learning models. Despite their prevalence, a comprehensive theoretical understanding of these methods remains elusive, particularly regarding their performance across different datasets and model architectures. Input features play a crucial role in shaping model predictions and, consequently, their calibration. However, the interplay between feature quality and calibration performance has not been thoroughly investigated. In this work, we present a rigorous theoretical analysis of post-hoc calibration methods, focusing on Platt scaling and isotonic regression. We derive convergence guarantees, computational complexity bounds, and finite-sample performance metrics for these methods. Furthermore, we explore the impact of feature informativeness on calibration performance through controlled synthetic experiments. Our empirical evaluation spans a diverse set of real-world datasets and model architectures, demonstrating consistent improvements in calibration metrics across various scenarios. By examining calibration performance under varying feature conditions utilizing only informative features versus complete feature spaces including noise dimensions, we provide fundamental insights into the robustness and reliability of different calibration approaches. Our findings offer practical guidelines for selecting appropriate calibration methods based on dataset characteristics and computational constraints, bridging the gap between theoretical understanding and practical implementation in uncertainty quantification. Code and experimental data are available at: https://github.com/Ajwebdevs/calibration-analysis-experiments.

Prognostic information is essential for decision-making in breast cancer management. Recently trials have predominantly focused on genomic prognostication tools, even though clinicopathological prognostication is less costly and more widely accessible. Machine learning (ML), transfer learning and ensemble integration offer opportunities to build robust prognostication frameworks. We evaluate this potential to improve survival prognostication in breast cancer by comparing de-novo ML, transfer learning from a pre-trained prognostic tool and ensemble integration. Data from the MA.27 trial was used for model training, with external validation on the TEAM trial and a SEER cohort. Transfer learning was applied by fine-tuning the pre-trained prognostic tool PREDICT v3, de-novo ML included Random Survival Forests and Extreme Gradient Boosting, and ensemble integration was realized through a weighted sum of model predictions. Transfer learning, de-novo RSF, and ensemble integration improved calibration in MA.27 over the pre-trained model (ICI reduced from 0.042 in PREDICT v3 to <=0.007) while discrimination remained comparable (AUC increased from 0.738 in PREDICT v3 to 0.744-0.799). Invalid PREDICT v3 predictions were observed in 23.8-25.8% of MA.27 individuals due to missing information. In contrast, ML models and ensemble integration could predict survival regardless of missing information. Across all models, patient age, nodal status, pathological grading and tumor size had the highest SHAP values, indicating their importance for survival prognostication. External validation in SEER, but not in TEAM, confirmed the benefits of transfer learning, RSF and ensemble integration. This study demonstrates that transfer learning, de-novo RSF, and ensemble integration can improve prognostication in situations where relevant information for PREDICT v3 is lacking or where a dataset shift is likely.

Mafia is a social deduction game where informed mafia compete against uninformed townsfolk. Its asymmetry of information and reliance on theory-of-mind reasoning mirror real-world multi-agent scenarios, making it a useful testbed for evaluating the social intelligence of large language models (LLMs). To support a systematic study, we introduce Mini-Mafia: a simplified four-player variant with one mafioso, one detective, and two villagers. We set the mafioso to kill a villager and the detective to investigate the mafioso during the night, reducing the game to a single day phase of discussion and voting. This setup isolates three interactive capabilities through role-specific win conditions: the mafioso must deceive, the villagers must detect deception, and the detective must effectively disclose information. To measure these skills, we have LLMs play against each other, creating the Mini-Mafia Benchmark: a two-stage framework that first estimates win rates within fixed opponent configurations, then aggregates performance across them using standardized scoring. Built entirely from model interactions without external data, the benchmark evolves as new models are introduced, with each one serving both as a new opponent and as a subject of evaluation. Our experiments reveal counterintuitive results, including cases where smaller models outperform larger ones. Beyond benchmarking, Mini-Mafia enables quantitative study of emergent multi-agent dynamics such as name bias and last-speaker advantage. It also contributes to AI safety by generating training data for deception detectors and by tracking models' deception capabilities against human baselines.

Machine learning (ML) is often viewed as a powerful data analysis tool that is easy to learn because of its black-box nature. Yet this very nature also makes it difficult to quantify confidence in predictions extracted from ML models, and more fundamentally, to understand how such models are mathematical abstractions of training data. The goal of this paper is to unravel these issues and their connections to uncertainty quantification (UQ) by pursuing a line of reasoning motivated by diagnostics. In such settings, prevalence - i.e. the fraction of elements in class - is often of inherent interest. Here we analyze the many interpretations of prevalence to derive a level-set theory of classification, which shows that certain types of self-consistent ML models are equivalent to class-conditional probability distributions. We begin by studying the properties of binary Bayes optimal classifiers, recognizing that their boundary sets can be reinterpreted as level-sets of pairwise density ratios. By parameterizing Bayes classifiers in terms of the prevalence, we then show that they satisfy important monotonicity and class-switching properties that can be used to deduce the density ratios without direct access to the boundary sets. Moreover, this information is sufficient for tasks such as constructing the multiclass Bayes-optimal classifier and estimating inherent uncertainty in the class assignments. In the multiclass case, we use these results to deduce normalization and self-consistency conditions, the latter being equivalent to the law of total probability for classifiers. We also show that these are necessary conditions for arbitrary ML models to have valid probabilistic interpretations. Throughout we demonstrate how this analysis informs the broader task of UQ for ML via an uncertainty propagation framework.

Faithful free-text explanations are important to ensure transparency in high-stakes AI decision-making contexts, but they are challenging to generate by language models and assess by humans. In this paper, we present a measure for Prediction-EXplanation (PEX) consistency, by extending the concept of weight of evidence. This measure quantifies how much a free-text explanation supports or opposes a prediction, serving as an important aspect of explanation faithfulness. Our analysis reveals that more than 62% explanations generated by large language models lack this consistency. We show that applying direct preference optimization improves the consistency of generated explanations across three model families, with improvement ranging from 43.1% to 292.3%. Furthermore, we demonstrate that optimizing this consistency measure can improve explanation faithfulness by up to 9.7%.

Manual modeling in Constraint Programming is a substantial bottleneck, which Constraint Acquisition (CA) aims to automate. However, passive CA methods are prone to over-fitting, often learning models that include spurious global constraints when trained on limited data, while purely active methods can be query-intensive. We introduce a hybrid CA framework specifically designed to address the challenge of over-fitting in CA. Our approach integrates passive learning for initial candidate generation, a query-driven interactive refinement phase that utilizes probabilistic confidence scores (initialized by machine learning priors) to systematically identify over-fitted constraints, and a specialized subset exploration mechanism to recover valid substructures from rejected candidates. A final active learning phase ensures model completeness. Extensive experiments on diverse benchmarks demonstrate that our interactive refinement phase is crucial for achieving high target model coverage and overall model accuracy from limited examples, doing so with manageable query complexity. This framework represents a substantial advancement towards robust and practical constraint acquisition in data-limited scenarios.

Neural Information Processing Systems http://nips.cc/