LOCUS: Task-Aware Low-Rank Post-Training for Token-Efficient Language Generation

A new arXiv paper examines how the way post-training updates are parameterized influences the length of text that large language models produce. The authors propose LOCUS, a task-aware low-rank adaptation method intended to curb the verbosity that standard preference alignment tends to introduce without sacrificing usefulness. The work targets serving costs, which grow with output length.