A team from Stanford University and NVIDIA Research put out a project page this week for a model that never writes a sentence of text. It only scores options. Given a situation and a list of possible next actions, the model, called a Contrastive Language Model or CLM, picks whichever action’s internal representation lines up most closely with the situation’s own representation, the way a search engine matches a query to a result rather than writing one from scratch.
That distinction matters because checking an AI agent’s next step is usually almost as expensive as generating it in the first place. Most coding and computer-use agents today ask a full language model to look at several candidate moves and judge which one is best, which means every safety check or quality filter adds a second expensive inference call on top of the first.
The Stanford and NVIDIA team, describing their own unpublished results, report that an 8 billion parameter CLM matches the picks of Jev, the model it is tested against, on tasks that span using a computer, playing games and calling tools, with the speed gap reaching a factor of nine in CLM’s favor. The biggest gains show up when there are many candidate actions to rank, or when the same handful of actions repeat across many situations. At around 1,000 candidates, the team measured CLM at 13 times faster than Jev on an Nvidia H100 chip. These are the authors’ own benchmarks, run on their own hardware, and have not been independently verified.
The team also tested CLM as a verifier rather than a generator: instead of writing code itself, it picks the best solution out of several candidates that another model produces first. For the DeepSWE benchmark, those candidates came from Anthropic’s Opus 5. For Terminal-Bench 2.1, they came from Anthropic’s Fable 5. In that role, and only after additional fine-tuning on top of the base model, CLM reached 81.6 percent on DeepSWE and 87.6 percent on Terminal-Bench 2.1, which the authors call state of the art. Jev, tested the same way, scored below the baseline of picking a candidate at random. The test sets behind those numbers are narrow: DeepSWE contributed 38 held-back tasks, and Terminal-Bench 2.1 supplied another 30, so the state-of-the-art claim rests on a sample too small to treat as settled.
The architecture is built to be cheap to run. Each of the two encoders, one for the situation and one for the action, is a frozen language model with a small trainable layer bolted on top, so only about 20 million parameters actually get trained. Because situations and actions are scored separately, a repeated action never needs to be re-encoded, which is why the speedup is largest in settings such as a video game where the same four moves recur every frame. The team says one full pass through their Nemotron question-and-answer set finishes in roughly sixty minutes on a single consumer-grade RTX 4090 card, a small fraction of what training a comparable general-purpose model would cost.
Nvidia’s own inference stack already leans on smaller, specialized models to cut the cost of running larger ones, and CLM reads as an extension of that strategy into agent verification specifically. If the speed and accuracy numbers hold up outside the authors’ own test sets, it gives teams building coding or computer-use agents a cheaper way to filter candidate actions without calling a full model for every check, though the released benchmarks so far cover only two narrow, small held-out sets. A multimodal version, CLM-35B, is still in training, with the team saying a release is planned for next month.
Teams building agent pipelines that currently spend a full model call verifying each step should watch for that CLM-35B release and an independent benchmark before swapping in a verifier this small.
Reported from the Stanford University and NVIDIA Research project page, published 23 September 2026.
