Pandora's Regret: A Proper Scoring Rule for Evaluating Sequential Search

Modern tabular foundation models such as TabPFN and TabICL naturally produce full predictive distributions, while the benchmarks used to evaluate them (TabArena, TALENT, and others) still rely almost exclusively on point-estimate metrics (RMSE, $R^2$). This mismatch implicitly rewards machine learning models or pipelines that elicit a good conditional mean while ignoring the quality of the predictive distribution. We make the case for using proper scoring rules for training, fine-tuning, and benchmarking (ranking) of tabular foundation models. Although all strictly proper scoring rules are theoretically equivalent at the population level, they may differ on finite data: We demonstrate analytically and empirically that different scoring rules can induce different inductive biases during finite-sample optimization, leading to different model performance. We validate this finding by running fine-tuning experiments with TabPFN and TabICL using different scoring rules for various data sets, revealing non-trivial interactions between training objectives and evaluation metrics. Our results show that practitioners can adapt tabular foundation models to task-specific scoring objectives, and that the choice of scoring rule can influence model behavior in practice.

Language-model judges and reward models enable scalable supervision, but finite optimization can exploit evaluator errors rather than improve response quality. We characterize this failure through the covariance geometry of evaluator ensembles. For calibrated judges, the ensemble mean retains common-mode error along the all-ones direction, whereas cross-judge disagreement captures only orthogonal error. Consequently, disagreement can be high despite robust aggregation, or low while shared response-dependent errors persist. We prove that common-mode error is not identifiable from internal judge scores alone. Under a joint sub-Gaussian model, we bound best-of-K selection overstatement and target-quality regret, extending the guarantees to predictably adaptive search under conditional calibration. The resulting search terms scale as the square root of log K and are asymptotically tight for Gaussian projected errors. We further show that noisy quality proxies introduce artificial rank-one covariance without changing disagreement, and propose a bounded two-anchor Bernstein certificate for finite-search error and regret. Fixed-seed Gaussian stress tests over 120 (J, rho, K) configurations and real-model audits validate the theory while revealing the limits of disagreement-based diagnostics under increasing search pressure.

Standard classification treats all errors equally, but in applications such as content moderation and medical screening, mistakes on clear-cut cases are more costly than errors on ambiguous ones. From a contextual bandit framework, we propose normalized excess cost (NEC), a metric that weighs classification errors by per-example costs and reduces to standard error rate when costs are uniform. Costs can derive from annotator vote margins, distance from decision thresholds, or confidence ratings. Across text, image, and tabular benchmarks, we find that NEC is often substantially lower than error rate: models with 5% error rate can achieve 1.8% NEC, revealing that most mistakes concentrate on ambiguous, low-cost examples. We also find that incorporating costs into training via loss weighting, sampling strategies, or regression yields inconsistent benefits. Our framework provides a practical methodology for deriving and evaluating instance-level misclassification costs, even if cost-sensitive training offers limited benefit.