Co-LMLM: Continuous-Query Limited Memory Language Models.

arXiv:2607.07707v2 Announce Type: replace-cross Abstract: Externalizing knowledge in LLM pre-training is a promising avenue to achieve higher performance at smaller scales, control knowledge use, and overall increase model transparency. We propose continuous-q