MoREGT pairs expert routing with an LLM agent for cross-database prediction
A new arXiv paper proposes a graph transformer that routes database rows and cells to specialized experts, plus a label-blind agent that adds derived columns to target databases.
In plain words
Predictive models usually must be retrained for each database because every platform stores data differently. The paper's authors built a model that treats rows and cells as tokens and sends them to specialized experts based on column metadata. A separate language-model agent also creates extra useful columns from the target database's schema, without seeing any labels.
What was announced
The authors say their goal is to show that a stronger architecture, combined with target-database features, can match gains that previously required much larger pretraining corpora. MoREGT is a mixture-of-relational-experts graph transformer: it represents rows and cells as tokens, feeds row log-ages and foreign-key relationship names into attention, and assigns rows and cells to experts using column metadata instead of values. A companion LLM agent reads the target database's schema and column statistics, then adds validated single-row derived columns while remaining blind to labels and model scores.
On 21 RelBench tasks, with leave-one-database-out pretraining across six databases, the reported backbone gains over a Relational Transformer pretrained identically by 2.25 mean AUROC points and 5.83 mean nMAE points. The agent's columns lift mean AUROC from 78.45 to 79.63 and cut mean nMAE from 27.91 to 26.98, which the authors report as the best means among compared methods, including RT-J, which pretrains on roughly 400 times as many databases. These are the paper's own reported results and have not been independently confirmed.
What it does not show
- Evidence is limited to the paper's abstract; experimental setup, baselines and code details are not visible in the supplied passages.
- All performance figures are author-reported on RelBench tasks and no independent verification is available.
For security and IT teams
- Scope
- The agent works only from schema and column statistics of the target database and never accesses labels or model scores, per the paper.
Sources
Source-support review completed 17 Nov 2026. This story was drafted with AI assistance and checked against retained source passages. Automated checking is not independent verification.
Coverage: arXiv: abstract.
Updates and corrections
No corrections so far.
First published 17 Nov 2026
New to the subject? AI from the Ground Up starts from the beginning.