Yifan-Song793/ETO

Implements iterative policy learning through contrastive trajectory pairs using DPO loss on failure-success examples, rather than relying solely on expert demonstrations. Provides integrated environments for WebShop, ScienceWorld, and ALFWorld, with a FastChat-based training pipeline supporting parallel exploration and multi-round optimization. Demonstrates significant generalization gains (22% improvement on out-of-distribution tasks) and maintains task-solving efficiency through fewer action steps.

159 stars. No commits in the last 6 months.