Stanford and Nvidia researchers have released Contrastive Language Models (CLM-8B), an open model that turns agent decision-making into a matching problem instead of a token-generation task. The model caches reusable action representations and ran up to 9x faster than TypeSafe's Jev in zero-shot tests across computer use, gaming and tool calling, while matching Jev's success rate on two game tasks.
The CLM approach targets a common inefficiency in AI agents: large language models often generate full token sequences even when the application only needs a bounded decision, such as picking a tool from a fixed list. CLM-8B encodes the current state and available actions into a shared embedding space, then selects the action that best matches the state. This sidesteps autoregressive output generation entirely.
Project lead Jacky Kwok of Stanford told VentureBeat the contrastive training objective shows specific advantages for decision tasks. "We found that the contrastive objective can be especially effective for decision-making tasks compared with standard supervised fine-tuning (SFT) using cross-entropy loss," he said. "We also found that initializing from a pretrained CLM checkpoint enables the model to adapt quickly to downstream/domain-specific tasks."
How the matching architecture works
CLM-8B uses a frozen Qwen3-8B backbone with separate projection heads for states and actions. During training with a method called InfoNCE, the model learns to pull correct state-action pairs closer together while pushing incorrect pairs apart. The training pipeline starts with roughly 60 million question-answer pairs for broad semantic matching, adds around 30 million synthetic hard negatives, and finishes with roughly one million agent trajectories to adapt the model to agent decisions.
At inference time, the model encodes the current state, compares that representation against cached action embeddings, and picks the closest match. If an application repeatedly chooses among the same actions - say, 50 approved IT support actions like resetting passwords or escalating tickets - CLM can encode those actions once and reuse the embeddings across requests. Only the changing state needs fresh encoding each time.
Performance trade-offs and speed gains
In zero-shot tests, CLM-8B matched Jev's success rate on two game tasks but gave up accuracy on two others. It scored 95.2% to Jev's 99.2% on the BFCL v4 tool-calling benchmark and completed 26 of 30 WikiRacing tasks to Jev's 30. The largest speedups appeared when the model could reuse actions across many states or choose from a large candidate set.
The researchers also tested CLM as a verifier for coding tasks. Larger models - Opus 5 for DeepSWE and Fable 5 for Terminal-Bench 2.1 - generated candidate solutions, and a fine-tuned CLM ranked and selected the best one. On held-out subsets, CLM reached 81.6% on 38 DeepSWE tasks and 87.6% on 30 Terminal-Bench 2.1 tasks, compared with 71.1% and 83.1% for Jev. CLM's verifier latency was 4.1-5.7x lower than Jev's.
Kwok described the division of labor this way: "Use large reasoning models to generate and reason, and use CLMs to cheaply select, verify, and monitor their outputs." He cautioned that CLM is not a replacement for general-purpose reasoning on tasks requiring open-ended problem solving. "I would not use a CLM as a replacement for a general-purpose reasoning model on tasks that require substantial open-ended reasoning, such as solving a math problem, generating a long-form response, or doing high-level planning," he said.
Where CLM fits among emerging decision models
CLM arrives alongside growing interest in what TypeSafe calls "System One" models - fast, structured decision systems distinct from slower reasoning models. TypeSafe's Jev, released earlier this month, uses Reinforcement Learning for Calibrated Decisions. Laya takes a different approach with smaller bidirectional encoders. CUA-S1 focuses narrowly on bounded actions for form-filling tasks.
The architectural difference matters most when both context and candidate actions are long. "A key difference between CLM and models like Jev or Laya is the architecture," Kwok said. Jev and Laya primarily support caching the state representation, while CLM's dual-encoder design lets it compute and cache state and action embeddings independently. "CLM is particularly well suited to applications with long context / reusable action spaces," he said.
The team released CLM-8B weights under Apache 2.0 alongside open-source code, a TypeSafe-compatible API, fine-tuning tools, and a playground for testing states and candidate rankings. A multimodal CLM-35B-A3B is now training with more data and compute, with a release planned for early October. The team is also adding substantially more agentic training data, with the goal of making the model easier to integrate into existing workflows.
Why this matters for product and engineering teams
CLM-8B changes the cost structure for multi-step agent workflows. One routing call through a generative model might add negligible latency, but agents often make dozens of routing, ranking and checking decisions before returning an answer. Those costs compound. By removing autoregressive generation from bounded decisions, CLM cuts per-decision latency and compute. The model also adapts to domain-specific tasks without fine-tuning the full 8-billion-parameter backbone - only the smaller projection heads need training. For teams building internal agents that choose repeatedly from a fixed toolset, the cached-action pattern offers a direct path to lower latency without swapping out the larger reasoning models that handle open-ended work.
Your membership also unlocks:
