CLM-8B 对比 Jev：System One 决策模型选型
访谈介绍斯坦福与英伟达的 System One 模型 CLM-8B，称其多数任务追平 Jev 且快至 9 倍，适合当推理模型的输出裁判。
AI agents and agentic workflows are getting more complex, but the primitives that support them have not caught up, especially on tasks that don't need a generative answer. This is the gap that System One models (or decision models) fill. Jev is the most publicized example but not the only game in town. CLM-8B by Stanford and Nvidia is a System One model based on Contrastive Language Modeling. CLM encodes and matches states and actions. It can take an input such as a sequence of text or code tokens and map them to one of several actions. One of the use cases I'm especially bullish on is using it as a judge for multiple outputs from reasoning models, which it can do at a very high speed/accuracy while cutting down compute costs significantly. In experiments, CLM ran up to 9x faster than Jev while matching it on most tasks. And it is composed of a frozen Qwen3-8B backbone and a small trainable MLP projection head, which means fine-tuning it is very cost-efficient. I spoke to project lead @jackyk02 on the vision behind CLM, the best applications, and the future directions.
这条是引用着另一条写的：x.com/venturebeat/status/2103578683197812905