New CLM model speeds AI agent decisions using contrastive learning, not generation

Researchers developed CLM, a contrastive language model designed specifically for AI agent decision-making rather than text generation. Unlike traditional LLMs that generate actions token-by-token, CLM scores state-action pairs directly through a contrastive learning objective. The 8-billion-parameter model reportedly achieves performance comparable to alternatives while reducing decision latency by up to nine times. Available as open-source software with Apache 2.0 licensing, it includes a Python package, API, and fine-tuning tutorials on Hugging Face. The approach aims to address the high cost and slow speed of using large generative models for every step in AI agent workflows.