Stanford Hazy Research has unveiled CLM-8B, an open-source contrastive language model designed for agentic systems. Unlike traditional generative models that produce text token by token, this 8B-parameter model selects actions by evaluating their compatibility with the current environment state.
The architecture is based on Qwen3-8B and utilizes a contrastive approach to decision-making. Instead of generating a response, the model computes a state vector f(s) and action vectors g(aᵢ), selecting the optimal action through the dot product of these vectors. This method effectively transforms the task into a classification problem.
The model is trained using the InfoNCE loss function, which pulls correct state-action pairs together while pushing incorrect ones apart. The probability of selecting a specific action is determined by the similarity between the state and action representations, adjusted by a temperature parameter τ.
This approach offers significant performance advantages when dealing with large sets of reusable actions. By pre-calculating and caching action vectors, the system avoids the computational overhead of sequential text generation, making it suitable for complex agentic environments.
