Effect of harness-design initialization on exploration and transfer

Determine whether initializing reinforcement learning with a model that has more prior knowledge of harness design and revision improves exploration and transfer in harness learning.

Background

The Reasoning Gym experiments initialize the proposer using supervision from a single 35B teacher, whereas the multi-hop question-answering experiments apply reinforcement learning directly to the base proposer without teacher-generated demonstrations. This leaves open whether a proposer with stronger prior knowledge of harness construction and revision would explore more effective revisions.

The question concerns both exploration during training and transfer to tasks not encountered during training. The paper explicitly identifies this as an unresolved question rather than reporting an answer.