A new paper on arXiv, titled "Language Models are Lazy Poets: Creative Mode Collapse at Scale," warns that aligned large language models, while capable of handling creative assignments, tend to produce outputs that are strikingly similar to one another. The authors frame this as a form of mode collapse, a known failure mode in generative models where variety is lost.
The abstract indicates that the problem persists even as models scale, and that while various solutions have been suggested in prior work, a comprehensive understanding remains incomplete. The paper appears to offer a large-scale investigation into the phenomenon, though the full details are not available in the abstract.
This work adds to a growing line of research questioning whether LLMs truly exhibit creativity or merely recombine familiar patterns. The "lazy poet" metaphor captures the concern: the models take the easiest path, producing competent but repetitive verse rather than genuinely novel expression.