Contrastive Language Models: A New Paradigm for Fast Decision-Making

Authors: Jacky Kwok$^{\dagger}$, Hangoo Kang, Tarun Suresh, Jon Saad-Falcon, Marco Pavone, Christopher Ré, Azalia Mirhoseini Affiliations: :stanford: Stanford University & :nvidia: NVIDIA Research Date: 🗓️ September 23, 2026 Lead: $^{\dagger}$ Project Lead

Core Thesis: We present Contrastive Language Models (CLMs), a "System One" architecture designed for rapid, generalizable decision-making. By utilizing a contrastive learning objective, CLMs create a direct mapping between environmental states and the optimal actions to take.

---

🚀 The CLM-8B Blueprint

The researchers have released CLM-8B, a model developed through a rigorous three-stage data pipeline:

• Pre-training: 60 million Nemotron QA pairs.
• Mid-training: 30 million synthetic "hard negatives" to refine discrimination.
• Post-training: 1 million high-quality agentic trajectories.

Performance Benchmarks

CLM-8B matches the capabilities of the Jev model across several domains—including tool-calling, gaming, and general computer use—but does so with significantly higher efficiency.

---

🛠️ Technical Architecture & Mechanism

Unlike traditional models that might rely on slow, token-by-token autoregressive generation, CLM utilizes a dual-encoder strategy.

The Contrastive Process

The model employs a state encoder and an action encoder. Using the InfoNCE objective, the system ensures that the embedding of a given state is mathematically "pulled" toward the correct action and "pushed" away from incorrect ones.

Deployment Logic:

1. Input the current state.
2. Generate embeddings for a set of candidate_actions.
3. Calculate the alignment (similarity) between the state and each action.
4. Select the action with the maximum score.

Infrastructure Innovation

To maximize throughput, the team implemented a disaggregated infrastructure. By separating the state and action components, the system can independently cache and reuse state_embeddings and action_embeddings, drastically reducing redundant computation.

---

📈 Scaling Laws & SOTA Results

The team discovered that CLMs follow predictable scaling trajectories. The test contrastive loss ($\mathcal{L}$) behaves as a power law relative to the following variables:

• Training Compute ($C$)
• Model Parameter Count ($N$)
• Dataset Volume ($D$)

$\mathcal{L} \propto C^{-\alpha} \cdot N^{-\beta} \cdot D^{-\gamma}$

Reward Model Excellence

When repurposed as a trajectory reward model, CLM achieves new State-of-the-Art (SOTA) results on complex coding benchmarks:

• DeepSWE: 81.6%
• Terminal Bench 2.1: 87.6%

---

🎮 Real-World Application

The efficiency gains are most evident in scenarios with massive action spaces or repetitive action patterns.

Figure 1: CLM-8B demonstrating zero-shot proficiency in Super Mario and WikiRacing.

Key Takeaway: The speedup is most dramatic in tasks like WikiRacing (large candidate sets) or the T-Rex Game (high action reuse).

Get Started: Check out the implementation on GitHub.