Stanford and NVIDIA's Contrastative Language Models match agent baselines at 9x lower latency

Researchers from Stanford University and NVIDIA Research have introduced Contrastive Language Models (CLMs), a new class of model designed for fast, generalizable decision-making in AI agents. According to the project page, posted on September 23, 2026 and surfaced on the Hacker News front page, the flagship CLM-8B model performs on par with the Jev baseline across computer-use, gaming and tool-calling tasks while achieving up to 9x lower latency.

A System One model for agents

The research team, led by Jacky Kwok and including co-authors such as Marco Pavone, Christopher Ré and Azalia Mirhoseini, positions CLMs as "System One" models. The framing borrows from psychology's dual-process theory: where large language models act as deliberate, token-generating "System Two" reasoners when an agent decides its next step, a CLM is built to make quick, intuitive decisions without generating text at decision time. The project page accompanies the release with interactive demos, including Dino Run, Super Mario and WikiRacing.

How the model chooses actions

A CLM trains two separate encoders, one for states and one for actions, using the InfoNCE contrastive objective. During training, a state's embedding is pulled toward the embedding of the action that was actually taken next and pushed away from all other actions. At deployment, the model receives the current state plus a set of candidate actions, scores each action by how well its embedding aligns with the state embedding, and selects the highest-scoring option. The two encoders effectively work as a zero-shot action classifier, and because no text generation is involved, the decision step is computationally cheap.

Training recipe and results

CLM-8B was built in three stages: pre-training on 60 million Nemotron question-and-answer pairs, mid-training on 30 million synthetic hard negatives, and post-training on 1 million agentic trajectories. In zero-shot evaluation across computer-use, gaming and tool-calling tasks, the researchers report performance on par with Jev while running up to nine times faster. The speedups were largest in settings with many candidate actions, such as WikiRacing, or where actions are frequently reused across states, such as the T-Rex game. With lightweight fine-tuning, the team says CLM-8B also sets new state-of-the-art results on agentic coding benchmarks, reaching 81.6 percent on DeepSWE and 87.6 percent on Terminal Bench 2.1.

Infrastructure and scaling laws

The team also describes an ultra-efficient training and serving setup that disaggregates states from actions, allowing their embeddings to be cached and reused independently. Beyond the model itself, the researchers establish scaling laws for CLMs, showing that test contrastive loss decreases predictably as a power law in training compute, model size and dataset size. The project's code is available on GitHub, according to the page.

Why it matters

Most AI agents today decide what to do next by prompting a large general-purpose language model and waiting for it to generate a response, which makes every step slow and costly. CLMs reframe action selection as an embedding-matching problem, closer to retrieval than generation, which could make agents viable in latency-sensitive settings such as games, UI automation and real-time control, or let cheap System One models handle routine steps while larger reasoning models are reserved for hard ones. The reported power-law scaling also suggests that improvements can be planned rather than discovered by trial.

The results are self-reported on the project's own page, and it does not explain what the Jev baseline is or how the evaluations were run, so independent verification will be needed before the numbers are settled. Even so, the combination of a new architecture, a disclosed training recipe and scaling analysis makes this a concrete attempt to make agent decision-making both fast and generalizable.