CLM 对比 Jev：快 9 倍的 System One 模型
斯坦福 CLM-8B 用对比学习替代 Jev 的 RLCD，在工具调用、游戏、computer-use 上以更低延迟追平 Jev 精度。
If you're building your own harness, watch this new wave of System One models First Jev. Now Contrastive Language Model, or CLM. CLM runs up to 9x faster than Jev. It also looks like a better verifier, especially on long-horizon tasks. How they differ: CLM is trained with contrastive learning. Jev is trained with Reinforcement Learning for Calibrated Decisions (RLCD). Jev takes a situation plus pre-defined questions and returns typed decisions with probabilities attached. CLM encodes the situation and possible actions into embeddings, compares their similarity, then ranks or picks the best match. The numbers are genuinely wild. Released as CLM-8B by Stanford researcher Jacky Kwok: on tool-calling, gaming, and computer-use tasks, it matches Jev's accuracy at up to 9x lower latency. After light fine-tuning as a verifier, it hits 81.6% on DeepSWE and 87.6% on Terminal-Bench 2.1, both new state of the art, while running 4-6x faster than Jev in that role. One sharper claim from the paper: in these long-horizon setups, Jev actually scored below the random-pick baseline as a verifier. CLM didn't. The bigger point here isn't which one wins. It's that there's more than one way to solve this problem, and that's genuinely interesting to watch play out. Code and weights are out under Apache 2.0.