Stanford's CLM-8B Speeds Up AI Agent Decisions Ninefold

Stanford's Scaling Intelligence Lab, in collaboration with Hazy Research, has released the repository, weights, serving infrastructure, and client libraries for CLM-8B under the Apache 2.0 license. This Contrastive Language Model is designed as a fast, single-pass decision mechanism for AI agents. Instead of generating text, CLM-8B scores a given state against a finite set of candidate actions in an embedding space, selecting the action with the highest dot product. This approach targets agent steps with bounded choices, such as tool routing, verification, and computer use.

The model architecture pairs a frozen Qwen3-8B backbone with separate 20-million-parameter projection heads for states and actions. These heads are trained using a bidirectional InfoNCE loss objective over matched state-action pairs, which increases similarity for correct pairs while decreasing it for others. By disaggregating states and actions, the model allows action embeddings to be cached once and reused across different states. This design achieves up to a ninefold inference speedup compared to Jev, the researchers' generative-verifier baseline, while maintaining comparable accuracy on tool calling, gaming, and computer-use tasks.

After task-specific fine-tuning, CLM-8B achieved state-of-the-art results as a best-of-N verifier, scoring 81.6% on a 38-task DeepSWE subset and 87.6% on a 30-task Terminal-Bench 2.1 subset. For practitioners, this means agent loops can run significantly faster because scoring additional candidates becomes a cheap vector operation rather than an expensive generative pass. Looking ahead, the researchers plan to release a larger, multimodal CLM-35B checkpoint on Hugging Face in early October.