Deciding Without Generating: A Survey of Typed Decision Models, from Jev to Open Decision LLMs

Many applications of large language models need a decision rather than free text. Examples include assigning a category, giving a score, answering yes or no, or choosing the next action. Prior work has used language models for such decisions in several ways, including zero-shot classification, reward models and verifiers, LLM-as-a-Judge, safety classifiers, and model cascades. In most of these approaches, the model generates text, and the decision must then be extracted from it. Recently, a class of models has emerged that answers typed questions directly, without generating text. Given an input and a set of questions with predefined answer types, such a model returns, for each question, an answer in the declared format together with probabilities over the possible answers, and for some question types a confidence score. We call these models typed decision models. The commercial model Jev made this design widely known, and it was soon followed by many open models and by similar services from other providers. Research on typed decision models is growing quickly but is scattered across many communities. This survey reviews the area. We trace its roots in earlier work and give a formal description of typed decision models whose outputs depend on the question type. We analyze how these models are combined with stronger fallback models, and show that such a combination improves accuracy only if the fallback corrects more errors than it introduces on the inputs passed to it. We then review model designs, training methods, sources of efficiency, applications, benchmarks, and known failure modes. In the reported settings, typed decision models are often much faster and cheaper than generative models, and they are often competitive on tasks where the needed evidence is stated in the input. They tend to perform worse on tasks that require multi-step reasoning and against supervised models trained on in-domain data. Their probabilities also depend on how a question is asked, and they can be misled by option names and by untrusted input. We conclude with open problems for future research.