28 upvotes
Harness-S picks sibling fixes to train more general agents
What is Harness-S: Learning from Sibling Harnesses for Generalizable Agent Training?
Harness-S watches a set of related “harness” modifications that each repair a specific failure of a base policy. During training it dynamically selects the most useful harness, letting the agent learn from a broader range of situations.
Who it is for
RL researchers and engineers dealing with brittle policies can use it to boost robustness without hand-crafting a single perfect fix.
How to try it
Open the paper page on HuggingFace for the code and demo notebooks