RT-J preprint reports database predictions with 10x fewer examples

Researchers report that RT-J, an 86M-parameter model for predicting outcomes from relational databases, matches strong feature engineering pipelines with 10x fewer examples on average in RelBench in-context regression tasks. The preprint describes predictions made using hundreds of examples or fewer. The researchers pretrained RT-J on 639 web-crawled databases with 6,023 LLM-generated forecasting tasks, using contexts of up to 8,192 database cells. A lightweight random-walk retriever selects examples, while context tuning and ensembling increase computation at prediction time. The reported comparison concerns regression tasks in RelBench; it does not establish the same reduction across all database prediction workloads.