LLM2Jev Extracts Calibrated Decisions from Next-Token Probabilities

LLM2Jev is an architecture-preserving framework that extracts categorical probability distributions from LLMs without free-form text generation. Testing on Qwen3.5-4B shows that modern models can function as calibrated decision models out-of-the-box, though fine-tuning with tree-factorized listwise loss further optimizes candidate selection.