Search papers, labs, and topics across Lattice.
Southern University of Science and Technology
1
0
3
10
LLMs can be coaxed into better alignment with human preferences by simply truncating training responses to equal lengths, forcing the model to focus on the crucial prefix tokens often overlooked by standard Direct Alignment Algorithms.
