Mohsen Bayati

A Few Teacher Steps Go a Long Way: Cost-Efficient On-Policy Data Augmentation for Agent Post-Training

Junze Ye, Jiayi Cheng, Miao Lu, Michal Mankowski, Jose Blanchet, Mohsen Bayati

For LLM agents, supervised fine-tuning is not only about teacher labels' quality, but also about which interaction contexts those labels condition on. Pure behavioral cloning uses full teacher demonstrations, creating a mismatch between teacher-induced contexts seen in training a…Read more