Stanford and Nvidia's CLM-8B speeds up agent decisions by matching state-action representations

Researchers from Stanford University and Nvidia have released Contrastive Language Models (CLM), a new approach for AI agent decision-making, along with the CLM-8B model. Instead of generating tokens step by step, CLM-8B encodes the current state and available actions, then selects the action whose representation best matches the state. Because action representations can be cached and reused, the design is particularly efficient for workflows that involve repeated decisions.

In zero-shot tests across computer use, gaming, and tool calling, CLM-8B ran up to 9x faster than TypeSafe's Jev model. It scored 95.2% on the BFCL v4 tool-calling benchmark, compared with 99.2% for Jev, and completed 26 of 30 WikiRacing tasks, while Jev completed all 30.

CLM-8B uses a frozen Qwen3-8B backbone with separate projection heads for states and actions. This architecture lets developers precompute and cache action embeddings when applications repeatedly choose from a fixed set of options, reducing per-request computation. The model supports bounded decision patterns including option selection, yes/no probability scoring, ordered-scale scoring, and ranking arbitrary candidates, making it applicable to tool routing, ticket triage, and retrieval shortlisting.

The team trained the model with the InfoNCE contrastive objective in three stages: large-scale semantic matching, synthetic hard negatives, and post-training on agent trajectories. CLM-8B weights are released under the Apache 2.0 license, along with source code, a TypeSafe-compatible API, and fine-tuning tools.