Given a relational database (RDB) storing heterogeneous tabular information, how can we predict missing (or future) values in some target column of interest? As candidate targets are numerous in enterprise settings, it is preferable to avoid learning a new model from scratch each time there is a new prediction task. Frozen foundation models (FMs) based on in-context learning (ICL) and RDB-specific pre-training provide a viable solution, but the ideal design still remains largely unresolved. Perhaps unexpectedly, preliminary evidence has recently shown that existing single-table FMs (aka TFMs), when combined with suitable parameter-free relational encoders, can be quite competitive in practice despite zero exposure to RDB pre-training or relational data of any kind. In this work we first empirically confirm this phenomenon with a more diverse array of experiments. We then turn to our main contribution, namely, surfacing multiple complementary explanations for why TFMs behave so effectively. These explanations ultimately hinge on innate synergies whereby parameter-free encoders map relational effects to fixed-length embeddings that loosely mimic distributions for which TFMs were originally designed; in so doing, desirable single-table properties transfer to relational settings (e.g., convergent posterior predictive distributions). Collectively, our analysis forges new pathways for improving RDB FMs, particularly when we prioritize interpretability and inevitable integration within agentic systems.
Figures & tables
Figure 1: A TFM-based approach outperforms other RDB FMs on each benchmark.
Figure 2: Classification/Regression performance on RDBs induced from single-table temporal data.
Figure 3: Left : Prediction quality over synthetic RDBs formed via 4 different interaction types in child table \mbox{\boldmathT}^{2} . Colors denote models; line styles distinguish the 4 RDB versions. Results are shown as the width of \mbox{\boldmathT}^{2} increases by padding with uninformative columns. Right : Performance on quasi-synthetic data, whereby new ground-truth labels are generated via interactions between different column pairs (x-axis) within the real-world RDB rel-trial. Solid/dark bars show original results, while shaded/light bars represent revised performance after explicitly appending needed interaction terms to the appropriate table. The latter simulates a practical agentic scenario described in Section 7 .
Appendix figures & tables7 assets
Supplementary material from the paper’s appendix.
Appendix
Method
AB churn
OB ctr
RR cvr
SE upvote
SE churn
Avg AUROC ↑
Avg Rank ↓
Supervised
XGBoost
50.00
50.00
50.00
49.68
50.84
50.10
9.6
GNN
75.71
62.39
84.70
88.61
85.58
79.40
4.6
RelAgent
79.44
62.63
86.35
89.45
89.03
81.38
2.0
Foundation model
Closed-Source
Appendix
Table 1: Results on 4DBInfer binary classification tasks (AUC ↑ ). Bold = best per column, underline = second best.
salt-item
salt-sales
Avg MRR
Avg Rank
Method
Plant
Shipping Point
Incoterm
Incoterm
Office
Group
Payment Terms
Shipping Condition
↑
↓
Supervised
GraphSAGE
0.99
0.98
0.69
0.62
1
0.16
0.37
0.57
0.67
3.88
LightGBM
0.61
0.28
0.73
0.73
0.99
0.02
0.10
0.51
0.50
4.75
Foundation model
Closed-Source
Appendix
Table 2: Results on the SALT benchmark (MRR ↑ ). Bold = best per column; underline = second best.
mimic
ratebeer
arxiv
Avg AUROC
Avg Rank
Method
stay
beer-churn
user-churn
dormant
citation
↑
↓
Supervised
LightGBM
51.81
76.21
83.92
75.79
71.21
71.79
5.80
GraphSAGE
55.01
78.67
94.27
80.51
82.50
78.19
3.60
Rel-Agent †
–
84.70
97.43
83.33
82.61
–
–
Foundation model
Appendix
Table 3: Results on Relbench v2 binary classification tasks (AUROC ↑ ). Bold = best per column; underline = second best.
avito
trial
amazon
stack
hm
Avg AUROC
Avg Rank
Method
click
visit
out
user
item
eng
badge
churn
↑
↓
Supervised
LightGBM
53.60
53.05
70.09
52.22
62.54
63.39
63.43
55.21
59.19
18.88
GraphSAGE
65.90
66.20
68.60
70.42
82.81
90.59
88.86
69.88
75.41
6.38
RelGT
68.30
66.78
68.61
70.39
82.55
90.53
86.32
69.27
75.34
6.81
RelGNN
68.23
66.18
71.24
70.99
82.64
90.75
88.98
70.93
76.24
4.12
Appendix
Table 4: Results on Relbench v1 binary classification tasks (AUROC ↑ ). Bold = best per column, underline = second best.
avito
trial
amazon
stack
hm
Geo Mean
Avg Rank
Method
ctr
adverse
success
user
item
votes
sales
↓
↓
Base
Global Median
0.043
57.533
0.462
16.783
64.234
0.068
0.076
1.303
14.57
Entity Median
0.046
57.930
0.441
17.423
66.436
0.069
0.078
1.329
15.57
Supervised
LightGBM
0.041
44.011
0.425
16.783
60.569
0.068
0.076
1.220
12.50
Appendix
Table 5: Results on Relbench-v1 regression tasks (MAE ↓ ). Bold = best per column; underline = second best.
Task
Origin
TabPFN-v3
RDBLearn
KumoRFM-2
RT-J
Classification (ROC-AUC ↑ )
acquire_valued_shoppers_challenge
BeyondArena
0.6245
0.6098
OOM
0.5797
anes_voting_2026
BeyondArena
0.8620
0.8260
OOM
0.5000
bank_marketing_temporal_month
OpenML
0.8856
0.8555
0.8568
0.3762
electricity_temporal
OpenML
0.9542
0.9098
0.8752
0.6176
hotel_booking_demand
BeyondArena
0.8059
0.7265
0.6645
0.5755
Appendix
Table 6: Full results corresponding to Figure 2 , where we examine RDBs induced from single tables. Classification reports ROC-AUC ( ↑ ); regression reports R2 ( ↑ ). Best result per row in bold ; second best underlined . Avg (excluding OOM) represents mean performance computed only over tasks/rows where every method reports a value; Avg (all) includes all tasks (KumoRFM-2 omitted due to OOM entries).
Dataset
Task
RelAgent (GPT-5.2)
RelAgent (Qwen3-32B)
RDBLearn
RDBLearn+Agent
rel-amazon
item-churn
0.8284
0.7992
0.8270
0.8273
user-churn
0.7078
0.6922
0.6973
0.6968
rel-avito
user-clicks
0.6836
0.6510
0.7050
0.6994
user-visits
0.6779
0.6462
0.6714
0.6661
rel-event
user-ignore
0.8725
0.7591
0.7446
0.7873
user-repeat
0.7820
0.7473
0.7646
0.7753
Appendix
Table 7: RDBLearn+Agent results on RelBench-v1 binary classification tasks (AUROC ↑ ). Bold = best per column, underline = second best. We emphasize that RelAgent is fully-supervised, training numerous models per task, while RDBLearn is an FM.
Institute for Artificial Intelligence, Peking University · University of Illinois at Urbana-Champaign · Institute of Computing Technology, Beijing University of Post and Telecommunication +1