AIM Tabular Index

AIM Tabular Index Methodology

The AIM Tabular Index combines AIMultiple's tabular models benchmark with two public leaderboards, TabArena and BeyondArena.

Scoring

Each benchmark rates models in Elo. TabArena sets the default RandomForest to 1000 Elo, BeyondArena sets the default XGBoost to 1000, and AIM-Tabular sets the average of its models to 1000. We shift each benchmark so that default XGBoost scores 1000, then average the shifted ratings to get the index. On any one benchmark, a 200-point lead predicts a win on about 3 of 4 datasets.

Intervals and rank

Each benchmark gives every model a 95% confidence interval, and we average these the same way as the scores. We order two models only when their intervals do not overlap. Otherwise, rank shows a range. For example, 1-4 means the model can place anywhere from first to fourth.

Configuration

The public benchmarks contribute their default-configuration scores. AIM-Tabular runs every model without tuning or fine-tuning. On TabArena, tuning lifts XGBoost 126 points above its default.

Speed and cost

AIM-Tabular reports seconds per fold (mean time to fit and predict one cross-validation fold) and GPU cost (total fold time at $5.98 per GPU-hour). Foundation models ran on one NVIDIA B200 GPU. Classical models ran on CPU and have no GPU cost.