RESEARCH PROJECT

DeepAnalyze: Scalable Training-Trajectory Synthesis

DeepAnalyze: Scalable Training-Trajectory Synthesis

ICML 2026

Project Lead · Synthetic data, quality filtering, curriculum learning, agentic RL

  • Developed data-grounded trajectory synthesis for DeepAnalyze-8B, combining teacher-distilled reasoning and environment-generated interactions in the DataScience-Instruct-500K training corpus.
  • Built a questioner–solver–inspector pipeline: generate tasks and acceptance criteria from real data, execute multi-turn solutions, and filter trajectories using interaction checks and resulting environment changes.
  • Implemented a curriculum from single-skill SFT to multi-skill cold-start training and GRPO, using approximately 470K, 20K, and 15K samples, respectively.
  • Evaluated on 12 benchmarks; achieved 38.88% on DABStep and 61.7% on DS-1000, versus 15.34% and 54.8% for the single-ability training variant.

ruc-datalab/DeepAnalyze — 4,662 GitHub stars, checked 2026-09-29.