LLM2Jev: LLMs Are Already Jev-Style Decision Models — When and How to Fine-Tune Them

简体中文

L1 mechanism diagnostic: joint candidate scoring and KL anchoring. No 4B checkpoint or JevBench run.

Paper information

Field	Details
Paper	arXiv v1
Institution	Microsoft, identified by first-author email.
First public release	2026-10-01 (arXiv v1)
Author code	No public author implementation found as of 2026-10-03.
Local adapter or method key	llm2jev
Local reproduction code	src/auto_research/foundation_latest_20261003.py

Paper method and evidence

Use next-token probabilities over parenthesized numeric options to form a distribution over a finite candidate set. Fine-tune with a listwise decision loss and a KL term that anchors auxiliary outputs to the base model.

flowchart LR
  I[Numbered options] --> M[Joint listwise scoring]
  M --> A[KL anchor on auxiliary outputs]
  A --> O[Decision distribution]

Original paper figure

Figure 1 shows the decision-training and optimization path. From the original paper; copyright remains with the authors.

Core equation

\(p(o)\propto\exp\sum_t\log p([o]_t)\) and \(L=L_{list}+\lambda KL(p_{base}\|p_{aux})\).

Reported results

The authors report that untrained Qwen3.5-4B matched same-backbone community Jev models and that fine-tuning helped mainly on weaker tasks.

Local reproduction

Compare the mechanism-on path with its disabled/default state. The three-seed diagnostic is diagnostic_only=true; no formal relative improvement, capability ranking, 4B checkpoint run, or JevBench evaluation.