LLM-as-Jev: LLMs Are Already Jev-Style Decision Models -- When and How to Fine-Tune Them
The paper argues strong general-purpose LLMs can already act as categorical decision engines without extra training.
LLM-as-Jev reads decisions from next-token probabilities over bracketed numeric option IDs, avoiding free-form generation. In the authors’ tests, Qwen3.5-4B matched Jev-style models on the same backbone without training and handled arbitrary option counts, letter-logit alternatives, and image-based decisions. Fine-tuning helped weaker models and some cases like many-option intent routing, but gave less return on stronger backbones. The authors say KL anchoring limited damage to normal conversational generation, with LoRA performing best on capable models. HF Daily Papers' note
LLM-as-Jev reads decisions from next-token probabilities over bracketed numeric option IDs, avoiding free-form generation. In the authors’ tests, Qwen3.5-4B matched Jev-style models on the same backbone without training and handled arbitrary option counts, letter-logit alternatives, and image-based decisions. Fine-tuning helped weaker models and some cases like many-option intent routing, but gave less return on stronger backbones. The authors say KL anchoring limited damage to normal conversational generation, with LoRA performing best on capable models. HF Daily Papers' note
score 4