Search papers, labs, and topics across Lattice.
Incept Labs
2
0
5
DPO's gradient imbalance causes models to avoid bad answers instead of generating good ones – AdaDPO fixes this with a simple, drop-in loss modification that boosts performance.
SAE-guided model editing fails when treating SAEs as scalpels to filter task vectors, but succeeds when treating them as stethoscopes to select layers for unfiltered task vector injection.
