Language models tend to converge on the same kinds of answers to open-ended questions, even when a range of plausible responses exists. The new arXiv paper "Gap Detection: Mitigating Homogenization through Demand and Defensibility" takes this problem as its starting point. The authors argue that producing genuinely useful alternatives requires understanding both what the query is calling for and what existing answers already cover.
The proposed approach, gap detection, is described as a way to identify missing or underexplored response options. Rather than simply asking the model to be more diverse, it evaluates potential alternatives against two criteria: the demand expressed in the query and the defensibility of the new answer. That dual lens is meant to avoid generating variety for its own sake.
The abstract does not provide full implementation details, but the framing suggests a method that could be applied to evaluation or generation pipelines. The paper's contribution is conceptual: it reframes homogenization not as a lack of randomness, but as a failure to detect where the space of acceptable answers has room for more. As with any single-source summary, the findings are preliminary and not yet independently verified.