Large language models are usually thought of as text generators, but a new arXiv preprint argues they can also act as Jev-style decision models — systems that return categorical probability distributions over a fixed set of options, with no free-form text. That property lets software consume model outputs directly.
The paper, titled LLM2Jev, sets out to measure how far LLMs already exhibit this behaviour and to determine when fine-tuning is necessary. The abstract frames the work as an investigation into the extent of this capability and the conditions under which fine-tuning helps.
Since the full paper is not yet detailed in the abstract, the precise results and methodology remain unclear from the announcement alone. The claim itself is notable: it suggests that off-the-shelf LLMs may already be usable as decision engines, with fine-tuning required only in specific cases.