The authors propose LLM2Jev, which formats candidate options as bracketed numeric identifiers and extracts decisions from prefix-conditioned next-token probabilities, avoiding free-form text generation and architecture changes.
General-purpose LLMs already work as Jev-style decision models out of the box
The authors show that training-free next-token probability scoring on a 4B backbone matches fine-tuned community Jev models, and that fine-tuning helps only in targeted settings.
Independent
Yinheng Li · Justin Wagle
Research Digest··2 min read
Yinheng Li and Justin Wagle present LLM2Jev, a framework that converts any causal LLM into a Jev-style decision model by scoring bracketed numeric option IDs from next-token probabilities.
§