Stanford and Nvidia Unveil Faster AI Decision-Making Model with CLM-8B

Researchers from Stanford and Nvidia have released Contrastive Language Models (CLM), a new approach to decision-making in AI systems. CLM is designed for applications where agents make repetitive decisions, such as choosing between tools or ranking candidate outputs.

In these cases, traditional language models often generate sequences of tokens, even though the application only needs a bounded decision. CLM addresses this issue by creating representations of the current state and available actions, then selecting the action that best matches the state.

CLM can cache reusable action representations and avoid recomputing them for every request, resulting in significant speedups when agents make repeated decisions or choose from a large set of candidates. In zero-shot tests across computer use, gaming, and tool calling, CLM-8B ran up to 9x faster than TypeSafe's Jev and matched its success rate on two game tasks.

CLM uses a contrastive training method called InfoNCE, which trains the model by giving it one correct state-action pair alongside several incorrect ones and rewarding it for assigning the highest similarity to the correct match. The researchers trained CLM in three stages: teaching broad semantic matching with question-answer pairs, adding synthetic hard negatives, and post-training on agent trajectories.

CLM has several advantages, including reducing latency in multi-step agentic tasks and adapting quickly to downstream/domain-specific tasks. It can also serve as a monitoring layer for agents, continuously scoring actions or trajectories produced by an agent and flagging unusual behavior.