CLM (Contrastive Language Model)

CLM (Contrastive Language Model) is a dual-encoder architecture for scoring or selecting an agent's next action. It is built to run far faster than a language model asked to judge each candidate step by generating text. Jacky Kwok and collaborators at Stanford's Scaling Intelligence Lab, working with Hazy Research, introduced it in a September 2026 release announced on X and detailed on a project blog and GitHub repository, with no peer-reviewed paper yet. The team frames CLM as a "System 1" counterpart to slower, generative verifiers: instead of writing out a judgment, it retrieves and ranks.

A state encoder and an action encoder train jointly with a contrastive objective from the InfoNCE family used in self-supervised representation learning. This pulls each state's embedding toward its true next action and away from unrelated ones. Because the two encoders are separate, state and action embeddings can be cached and reused independently, which cuts latency when a fixed set of candidate actions is scored repeatedly against a changing state.

CLM-8B, the team's released model, is pretrained on 60 million Nemotron question-answer pairs, mid-trained on 30 million synthetic hard negatives, and post-trained on 1 million agentic trajectories. The team reports it runs up to nine times faster than their earlier LLM-as-a-Verifier framework while matching its accuracy on computer-use, gaming, and tool-calling tasks. Lightweight fine-tuning, they report, brings new results on DeepSWE (81.6%) and Terminal-Bench 2.1 (87.6%). They also report that CLM's contrastive test loss falls as a predictable power law in training compute, model size, and dataset size.