Contrastive-LM unveils CLM-8B, a new open model outperforming jev in agent action scoring

Contrastive-LM has launched CLM-8B, an open-source model designed to score agent actions up to 9 times faster than TypeSafe AI's proprietary Jev system. CLM-8B operates by evaluating candidate actions against the current state, returning probabilities instead of generating text. It is compatible with TypeSafe's API and runs on a single NVIDIA GPU under Linux, using the vLLM framework. The model employs a bidirectional InfoNCE loss to train state and action encoders, each built on a frozen Qwen3-8B backbone with a trainable projection head. At inference, CLM scores actions using dot products of embeddings, with softmax generating the final distribution. On benchmark tasks like T-Rex, CLM matches Jev's performance, though it lags in tool calling and WikiRacing. Pre-training achieves 52.1% top-1 accuracy, rising to 69.2% mid-training. CLM runs 13 times faster than Jev on 1,000 candidates, with latency improvements on an H100. The research team claims new SOTA results, with CLM outperforming Jev in held-out benchmarks.