Kumorfm 2 Eval

---
name: kumorfm-2-eval
description: Evaluates the in-context learning capabilities of a relational foundation model on multi-table predictive tasks across diverse domains. It probes the model's ability to perform binary classification, multi-class classification, and regression directly on relational database structures without flattening or fine-tuning. Use when the user wants to benchmark on RelBenchV1, RelBenchV2, SALT, 4DBInfer, or asks about evaluating this task. Reports AUROC.
metadata:
  skill_kind: dataset_eval
  source_arxiv: 2604.12596
  bibtex_key: hudovernik2026kumorfm2
  confidence: high
---

# kumorfm-2-eval

> KumoRFM-2: Scaling Foundation Models for Relational Learning — Hudovernik et al. (2026) (arXiv:2604.12596, 2026)

## What this evaluates

Evaluates the in-context learning capabilities of a relational foundation model on multi-table predictive tasks across diverse domains. It probes the model's ability to perform binary classification, multi-class classification, and regression directly on relational database structures without flattening or fine-tuning.

## Datasets

- **RelBenchV1** — total ?; splits: train (-1), val (-1), test (-1)
- **RelBenchV2** — total ?; splits: train (-1), val (-1), test (-1)
- **SALT** — total ?; splits: train (-1), val (-1), test (-1)
- **4DBInfer** — total ?; splits: train (-1), val (-1), test (-1)

## Metrics

- `AUROC` **(primary)** — range: [0, 1]
  - Area Under the Receiver Operating Characteristic curve. Computed per binary classification task and averaged across tasks within each benchmark suite.
- `Accuracy / F1 / MAE` — range: [0, 1] | percent | other
  - Task-dependent metrics: accuracy or macro-F1 for multi-class classification, and mean absolute error for regression tasks. Averaged across tasks per suite.

## Input / output format

**Input**: Multi-table relational database schema and data. Context examples from training and validation splits (up to 10k total) are provided for in-context learning, along with the target table/query for prediction.

**Output**: Predicted labels or regression values for the target entity/table row, obtained via in-context learning without fine-tuning.

## Scoring recipe

```python
def compute_auroc(y_true, y_pred):
    fpr, tpr, _ = roc_curve(y_true, y_pred)
    return auc(fpr, tpr)

def compute_suite_score(results):
    return mean([compute_auroc(y_true, y_pred) for y_true, y_pred in results])
```

## Common pitfalls

- Using only the training split for context instead of including validation splits when provided, as the protocol explicitly states to use both.
- Exceeding the 10k context example limit, which violates the defined realistic in-context learning setting.
- Fine-tuning the model, as the evaluation strictly mandates zero-shot/few-shot in-context learning on the base model without further training.

## Evidence (verbatim from paper)

> We assess the in-context learning capabilities of KumoRFM-2 on four different benchmark suites: RelBenchV1, RelBenchV2, SALT, and 4DBInfer. None of these datasets were used during pre-training, which guarantees no leakage of information. ... We use at most 10k context examples to reflect realistic settings and ensure fair comparison with prior work. ... Table 3: Test results on the binary classification tasks in RelBenchV1. Higher is better (AUROC).

## Citation

```bibtex
@misc{hudovernik2026kumorfm2,
  title={KumoRFM-2: Scaling Foundation Models for Relational Learning},
  author={Hudovernik et al. (2026)},
  year={2026},
  note={arXiv:2604.12596}
}
```

- arXiv: 2604.12596