Knowledge distillation (KD) usually transfers capabilities from a stronger teacher model to a weaker student by training the student's parameters. That works well, but it ties the distilled knowledge to one particular architecture, and the abstract of a new arXiv paper suggests this is a limitation.
The paper, titled "Universal Textual Teaching for LLMs," appears to propose a different route. Based on the available abstract, the method avoids training student parameters entirely, which would mean the distilled knowledge is not locked to a specific model design. The title hints that the teaching is done through text itself, rather than through weight updates.
However, the source is only a truncated abstract, so we do not know how the method works in practice or how it compares to standard KD. The announcement cuts off mid-sentence, leaving the stated drawback incomplete. Until the full paper is available, any further claims would be speculation.