What it is
Builds synthetic data-analysis worlds by specifying generative mechanisms before files, then uses audited rubrics for benchmarking and RL training of data agents. Instead of mining tasks from existing files, it generates the hidden world first, then projects noisy files and audited rubrics from that truth. This gives checkable open-ended data-analysis tasks and RL rewards without private data or trajectory distillation.
Key results
- +20.90 on DABstep-Hard over Qwen3.6-35B-A3B
- +14.28 on DAComp over the base model
- +7.06 on DataWorld-Bench over the base model
Explicit generative worlds can supply scalable, verifiable supervision for training data-analysis agents on open-ended insight tasks.