Search papers, labs, and topics across Lattice.
1
0
3
An on-policy expert-correction pipeline is developed, automated by a meta-level MLE agent, that localizes the failing turn in the weaker model's own rollout and asks the expert to rewrite only that turn, which preserves the model's planning style and combines the gains of harness evolution and model adaptation.
