LLM2Jev: LLMs Are Already Jev-Style Decision Models -- When and How to Fine-Tune Them
The paper argues many LLMs can act as calibrated option selectors without new model architecture.
LLM2Jev reads next-token probabilities over bracketed numeric option IDs to produce categorical decisions directly. The authors report that Qwen3.5-4B can match Jev-style models on the same backbone without training, including arbitrary option counts and image-based decisions. Fine-tuning helps weaker models and some tasks, especially many-option intent routing, but has diminishing returns on stronger backbones. KL anchoring is used to preserve conversational generation behavior during fine-tuning. ArXiv · AI/CL/LG's note
LLM2Jev reads next-token probabilities over bracketed numeric option IDs to produce categorical decisions directly. The authors report that Qwen3.5-4B can match Jev-style models on the same backbone without training, including arbitrary option counts and image-based decisions. Fine-tuning helps weaker models and some tasks, especially many-option intent routing, but has diminishing returns on stronger backbones. KL anchoring is used to preserve conversational generation behavior during fine-tuning. ArXiv · AI/CL/LG's note
score 4