Search papers, labs, and topics across Lattice.
1
0
2
LLMs can retain 10x more of their original capabilities after fine-tuning, simply by using a dynamically adjusted "anchor" to constrain distributional drift during training.
