Lost in Aggregation: How Benchmarks Overlook Irreplaceable Model Strengths

Benchmarking is essential for tracking progress in machine learning, offering standard reference points to compare methods and guide model development. In tabular machine learning, where datasets vary widely in size, noise, and feature interactions, evaluation results are typically compressed into aggregate metrics like average ranks, win rates, or Elo ratings. While these aggregates help identify robust default models, they conflate qualitatively different dimensions of model behavior. This paper argues that evaluation should also incorporate the data-centric peak performance frontier to reveal which models are truly necessary for peak performance on particular datasets.

To move beyond simple averages, the authors propose evaluating models using a data-centric peak performance frontier defined by the best statistically supported performance achieved on each dataset. Using paired fold-wise differences and a Wilcoxon signed-rank test, model performance on any given dataset is categorized into four distinct roles: (1) irreplaceable, if the model uniquely achieves the unmatched peak performance; (2) sufficient, if it reaches the peak frontier but shares this status with at least one other model; (3) redundant, if it misses the frontier without experiencing decisive failure; and (4) fallible, if it performs significantly worse than typical competitors. This taxonomy disentangles unique dataset-specific strengths from general consistency.

Applying this framework to the TabArena benchmark reveals that standard leaderboard aggregation metrics are heavily correlated and measure a dominant dimension driven by consistency and failure avoidance. Consequently, models that perform decently across all tasks without ever achieving the best score on any single dataset are heavily rewarded, whereas models with unique dataset-specific strengths often appear mediocre under aggregation. Furthermore, strong aggregate performance does not guarantee peak frontier coverage; for instance, top-ranked default models may rarely reach the absolute peak performance frontier, while specialized models that fail on many tasks can still reveal valuable and unique inductive biases.