释放内在Jev:开放决策模型的简单方案
Releasing the Inner Jev: A Simple Recipe for Open Decision Models
AI总结:
本研究提出一种简单方案,通过LoRA微调开放指令微调语言模型,构建OpenDecisions数据集,使模型在决策任务上超越Jev,取得最优性能并发布相关资源。
摘要:
基于大语言模型(LLM)的决策系统,也被称为System One模型,近来颇受欢迎,最具代表性的是Jev(Almeida,2026)。给定带类型的输入状态、一个问题和一组选项,这类模型会在单次前向传播中返回每个选项的概率,无需生成文本。我们证明,现有的开放指令微调语言模型无需任何额外训练,就能通过从模型最后一个token位置的输出logits中读取字母(A、B、C等)形式的答案,达到与Jev相当的性能。此外,我们还展示了一种简单的监督微调方案可进一步提升其性能。我们评估了四个开放指令微调语言模型:Gemma 4 31B、Gemma 4 12B、Qwen3.6-27B和Qwen3.5-4B。我们使用LoRA和从语言模型头部切片初始化的决策头对每个模型进行微调。此外,我们推出了OpenDecisions,这是一个包含2352个问题的新型验证数据集,旨在反映JevBench(JevBench维护者,2026)的任务构成,并基于准确率和校准指标计算能力得分。未经微调时,Gemma 4 31B在OpenDecisions上得分为79.6,而Jev为80.1。经过我们的微调方案后,它在OpenDecisions上得分为84.4,在JevBench v1.7上得分为81.5,分别比Jev高出4.3和4.4个能力得分点,在OpenDecisions上创下新的最优性能。在一组公开决策数据集上,我们的模型达到了与Jev相当的准确率,且校准误差更低。我们的结果表明,开放语言模型已具备System One能力,通过简单微调即可在广泛的决策任务上超越Jev。我们发布了我们的模型、推理代码和OpenDecisions。
英文摘要:
LLM-based decision-making systems, also called System One models, have recently gained popularity, most notably through Jev (Almeida, 2026). Given a typed input state, a question, and a set of options, they return a calibrated probability for each option in a single forward pass, without text generation. We show that open instruction-tuned LLMs can achieve performance comparable to Jev without any additional training, by reading out answers from letters (A, B, C, etc.) in the model's output logits at the last token position. Next, we propose a simple supervised LoRA fine-tuning recipe that can improve their performance further. We evaluate four instruction-tuned LLMs: Gemma 4 31B, Gemma 4 12B, Qwen3.6-27B, and Qwen3.5-4B. Furthermore, we introduce OpenDecisions, a novel 2,352-question validation dataset designed to reflect the task composition of JevBench (JevBench maintainers, 2026) and which computes a Capability score based on accuracy and calibration. Without fine-tuning, Gemma 4 31B scores 79.6 on OpenDecisions, compared to Jev's 80.1. After fine-tuning, it scores 89.0, outperforming Jev by 8.9 Capability points and on par with the strongest open decision model we evaluated. The same recipe improves the three other models by 7.8 to 15.7 points. On a suite of public decision datasets, our models achieve accuracy and calibration comparable to Jev. Our results show that open LLMs already possess System One capabilities, and with simple fine-tuning they outperform Jev on a wide range of decision tasks. We release our models, training and serving code, OpenDecisions, and DecisionBank, our training dataset consisting of 261,561 decisions.
机构、总结、中文摘要均由AI分析而来,可能存在错误,请以论文正文为准。