Beyond Policy Optimization: A Data Curation Flywheel for Sparse-Reward Long-Horizon Planning

In three linesBPO, a three-stage framework (bootstrapping, extrapolation, refinement), creates a self-improving data flywheel to train robust reasoning models for long-horizon sparse-reward planning. Uses planning quaternions, long-short chain-of-thought fusion, and complexity-stratified curriculum learning. SOTA on ALFWorld, ScienceWorld, WebShop with significant token efficiency.

Summary generated by Claude — human-verified