RESEARCH PROJECT

DeepPrep: Verifiable Task Synthesis for Agent Training

VLDB 2026

Project Lead; First Author · Executable supervision, reversible corruption, multi-turn RL

Designed a synthesis method that converts SQL workloads into source tables, target specifications, and executable pipelines; adds reversible noise and verifies recovery to preserve task solvability.
Constructed 6,788 training tasks, with pipelines spanning 31 operator types and up to 28 steps; trained agents through operator-level SFT, verified reasoning trajectories, and GRPO with hybrid rewards.
Synthetic training data improved out-of-domain accuracy by 37.0 percentage points over Parrot training on Synth-Bird (Qwen3-8B, data ablation); tree-based execution supports backtracking and pipeline repair.