LLM-as-Jev: LLMs Are Already Jev-Style Decision Models -- When and How to Fine-Tune Them

LLM-as-Jev. LLMs Are Already Jev-Style Decision Models -- When and How to Fine-Tune Them. By Yinheng Li and one other. Published on the fourth of October. No code repository is linked to it.

A Jev-style decision model returns a probability distribution over a set of predefined options, with no free-form text, so that software can act on its output directly. There is nothing to parse and no sentence to interpret: the system receives a distribution and picks from it. The question here is how far general-purpose language models already do this out of the box, and when fine-tuning is really needed. The authors' answer: modern models are inherently effective decision models, and fine-tuning helps in targeted places rather than everywhere.

The tool for asking is LLM-as-Jev, a framework that leaves the model's architecture untouched. It reads calibrated decisions directly from next-token probabilities over numeric identifiers placed in brackets. It comes in two forms: an inference recipe that needs no training, and a fine-tuning objective. The objective optimises candidate selection with a tree-factorised listwise loss, while KL divergence penalties anchor auxiliary predictions to the base model. The first half asks whether the capability is already there. The second half shapes it further when it is not, while trying not to disturb everything else the model does.

The models tested are Qwen3.5-4B and Qwen3-0.6B. Without any training, the larger of the two matches community Jev-style models built on the same backbone, outperforms readouts based on letter logits, supports any number of options, and natively handles multimodal decisions over images. That is the basis for the title's claim. A general model, read the right way, already behaves as a decision model, without any change to how it was built.

Fine-tuning's benefits are targeted rather than universal. It substantially improves weaker models and specific tasks, with many-option intent routing as the example given, but its returns diminish for strong backbones. One safeguard matters for keeping a model useful elsewhere: the KL anchors prevent behavioural degradation in conversational text generation. Among the fine-tuning setups, LoRA delivers the strongest performance on capable models.

For anyone deciding whether to train, the guidance is conditional. A capable backbone may already be enough when read through its own next-token probabilities. Fine-tuning earns its place with weak models or demanding tasks like routing across many options, and with anchors in place so the model can still hold a conversation.