Contrastive Language Models Accelerate Agent Decision-Making at Low Latency 🔗
CLM-8B matches Jev performance on agentic tasks while running up to nine times faster on modest hardware
Contrastive Language Models (CLM) introduce a System One architecture designed for rapid, generalizable decision-making in AI agents. By training a contrastive objective that links states and actions, the model learns to score candidate responses efficiently without generative overhead.
The CLM-8B variant, served via a TypeSafe-compatible API, decouples state and action embeddings, enabling independent caching and reuse that slashes latency during inference.
Pre-trained on 60M Nemotron Q&A pairs, mid-trained on 30M synthetic hard negatives, and post-trained on 1M agentic trajectories, CLM-8B achieves parity with Jev across computer-use, gaming, and tool-calling benchmarks. Its lightweight fine-tuning capability sets new state-of-the-art results as a verifier, scoring 87.6% on Terminal-Bench 2.1 and 81.6% on DeepSWE. Developers interact through client.system_one(state, questions), supporting structured or plain dict inputs, with a built-in rank engine for best-of-N selection. A local web UI at http://localhost:8700/ visualizes answer distributions and shares ranked results via links.
Deployment is streamlined: clm-serve launches the service, with UI disabled via --no-ui for API-only use. Long sequences require proportional scaling of --max-model-len and --max-tokens, increasing GPU memory demands. The project targets builders integrating fast verification into agent loops, especially where speed and precision outweigh generative flexibility.
The catch: Despite strong benchmark scores, the model’s reliance on cached embeddings may limit adaptability in highly dynamic environments where state-action relationships shift rapidly, and long-context handling remains constrained by truncation beyond 2048 tokens without manual configuration increases.
Why this leads today Contrastive-LM/CLM introduces a new method for LLM decision-making with structured, probabilistic outputs, gaining notable traction among developers seeking reliable AI integration without retraining.
- Embed fast verifiers in agentic coding pipelines
- Power low-latency tool selection in interactive agents
- Accelerate decision loops in gaming AI agents
Source: Contrastive-LM/CLM — based on the project README.