Tag
这篇论文介绍了一种"typed decision"System 1 接口：通过一次前向传播直接从模型 logits 或小分类头输出固定选项的概率，用于替代昂贵的 frontier LLM 判官。作者部署了一个自托管的 rank-16 LoRA 判官来分流 14.8 万条 agent 编码会话记录，成本仅为前沿模型判官的约 1%；并在 JevBench 上用 35B 开源模型达到 79.3%，与商业参考系统（73.0%）无显著差异。
微软宣布推出新型决策模型 Microsoft-Decision-1，专为快速决策任务设计，在结构化决策任务上于延迟和质量方面超越 LLM 及其他决策模型，目前已在事件响应、质量控制和科学发现等内部场景中测试。
Phocinae-Largha-150M-v1 是一个 144.3M 参数的决策模型（基于 mmBERT-small），接受结构化状态和类型化的是非/选择/评分问题，单次前向推理即可输出带置信度的校准答案，在自建英文评测集上取得 0.906 准确率，延迟低至 GPU fp16 21ms。
awesome-jev 的定期巡检新增 3 条条目（555 → 558）：llama.cpp 新增 System One 决策模型运行时支持、Databricks 推出 ai_decide AI Function 作为结构化决策 API，以及一个基于 Jev 官方文档的实际 RAG 测试案例。
Shanghai AI lab releases Intern Decision, a family of small multimodal models (0.8B, 2B, 4B) for structured decisions under Apache 2.0, based on Qwen3.5.
SelfJev is an open 4B decisions model that you can self-host, allowing typed questions (yes/no, pick one, pick any, score) over any text with calibrated probabilities, plus Jev-compatibility via TypeSafe's SDK and fine-tuning on your own data.
The article explains how to use Jev, a model for structured decisions, as an efficient judge for evaluating AI agent responses, reducing latency and cost compared to traditional LLM judges.
Jev, a probabilistic decision model from TypeSafe AI, has become the fastest-adopted model in Vercel's AI Gateway history, offering significant speed and cost advantages for structured decision-making in software applications.