CLM: The New Wave of System One Models for Fast Verification

Pay attention to the new wave of "System One" models if you are building your own harness systems for AI.

First came Jev. Now, there is the Contrastive Language Model (CLM).

Key Differences and Advantages

• Speed: CLM operates 9 times faster than Jev.
• Verification Role: CLM appears to perform better as a verifier, especially in tasks with a long execution horizon where rapid assessment is crucial.

Technical Breakdown: CLM vs. Jev

The core difference lies in the training methodology:

1. CLM uses contrastive learning. It encodes the current situation and potential actions into embeddings, compares their similarity, and then ranks or selects the most suitable option.
2. Jev is trained using Reinforcement Learning for Calibrated Decisions (RLCD). It takes a situation along with predefined questions and returns typed decisions with associated probabilities.

Why It Matters

The essence is that there are multiple ways to solve this problem, and this diversity is genuinely interesting for developers. Having different architectures allows for optimizing the trade-off between reasoning depth and speed in agentic workflows.

The code and weights are published under the Apache 2.0 license:

• GitHub Repository
• HuggingFace Weights