研究库 论文知识库

内容库 / 论文

arXiv:2610.02076 · 工程化

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

LLM-as-Jev：LLM 已是 Jev 风格决策模型 —— 何时以及如何对其进行微调

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

类型：arxiv
标识：2610.02076
链接：https://arxiv.org/abs/2610.02076
主分类：engineering
形态：method
TLDR：Jev-style decision models return categorical probability distributions over predefined options without generating free-form text, enabling software systems to act on their outputs directly. In this work, we investigate the extent to which general-purpose LLMs already possess this capability out of the box, and when fine-tuning is actually necessary. We present LLM-as-Jev, an architecture-preserving framework that extracts calibrated decisions directly from next-token probabilities over bracketed numeric identifiers. LLM-as-Jev provides both a training-free inference recipe and a fine-tuning ob
副分类：llm-infra
待LLM分类：否
标题中文：LLM-as-Jev：LLM 已是 Jev 风格决策模型 —— 何时以及如何对其进行微调
TLDR中文：Jev 风格决策模型在预定义选项上返回类别概率分布，无需生成自由文本，使软件系统能直接基于其输出行动。本工作研究通用 LLM 在开箱即用情况下已具备该能力的程度，以及实际何时需要进行微调。我们提出 LLM-as-Jev，一种架构保持的框架，通过方括号数字标识符的下一 token 概率直接提取校准决策。LLM-as-Jev 既提供免训练推理方案，也提供微调目标……
来源文件：
/inbox/tom/_candidates/2026-10-07-agent-rag-longcontext-candidates.json

相关 · 按语义匹配

仓库 AIDASLab/Awesome-LLM-Values-and-Pluralistic-Alignment

笔记 Context Length Alone Hurts LLM Performance Despite Perfect Retrieval — 精读与批判

仓库 MARKTECHPOST-AI-MEDIA-INC/LLMs-Tutorials-Projects

笔记精读笔记 · Position：LLM Serving 需要数学优化与算法基础

笔记 flyP 精读｜Agent-as-a-Judge: LLM-as-a-Judge 范式演化的第一篇综述

仓库 NoeFlandre/distill-abm

arXiv

2610.02076

TLDR

Jev 风格决策模型在预定义选项上返回类别概率分布，无需生成自由文本，使软件系统能直接基于其输出行动。本工作研究通用 LLM 在开箱即用情况下已具备该能力的程度，以及实际何时需要进行微调。我们提出 LLM-as-Jev，一种架构保持的框架，通过方括号数字标识符的下一 token 概率直接提取校准决策。LLM-as-Jev 既提供免训练推理方案，也提供微调目标……Jev-style decision models return categorical probability distributions over predefined options without generating free-form text, enabling software systems to act on their outputs directly. In this work, we investigate the extent to which general-purpose LLMs already possess this capability out of the box, and when fine-tuning is actually necessary. We present LLM-as-Jev, an architecture-preserving framework that extracts calibrated decisions directly from next-token probabilities over bracketed numeric identifiers. LLM-as-Jev provides both a training-free inference recipe and a fine-tuning ob

相关解读

arXiv ↗ 中文全文