Stanford's CLM Turns Agent Decisions Into Vector Search 9x Faster
CLM-8B replaces slow generative verifiers with a dual-encoder contrastive model that ranks actions up to 9x faster than Jev.
- Stanford's Scaling Intelligence Lab released CLM-8B, a contrastive System One model, under Apache 2.0.
- Dual state and action encoders trained with bidirectional InfoNCE loss on a frozen Qwen3-8B backbone plus 20M-param heads.
- Up to 9x faster than Jev on tool calling, gaming, and computer-use tasks with comparable accuracy.
- Fine-tuned SOTA on DeepSWE (81.6%) and Terminal-Bench 2.1 (87.6%) as a best-of-N verifier.
- State-action disaggregation lets action embeddings be cached once and reused across states.
- Multimodal CLM-35B checkpoint on Hugging Face arriving in early October.
CLM turns agent decisions into vector search
Stanford’s Scaling Intelligence Lab and Hazy Research have released CLM’s repository, weights, serving infrastructure, and client libraries. Contrastive Language Models score a supplied state against a finite set of candidate actions in embedding space, then select the action with the highest dot product. The design targets tool routing, verification, computer use, and other agent steps whose output is a bounded choice.
The initial CLM-8B release uses a frozen Qwen3-8B backbone and two small projection heads. Its authors report task success comparable to Jev, their generative-verifier baseline, with up to a ninefold inference speedup on zero-shot evaluations. After task-specific fine-tuning, CLM reached 81.6% on a 38-task DeepSWE subset and 87.6% on a 30-task Terminal-Bench 2.1 subset. Those coding results measure selection among solutions generated by other models, and the small evaluation sets warrant caution.
One pass over a bounded choice
The authors describe CLM as a “System 1” model, using the term as shorthand for a fast, single-pass decision mechanism. Its output is a ranking over supplied candidates, which makes the model suitable for the classifiers, routers, and verifiers embedded inside agent loops.
The model checkpoint pairs the frozen Qwen3-8B backbone with separate 20-million-parameter projection heads for states and actions. Training updates these heads while leaving the language model fixed. At inference time, each fresh state or action is encoded once, projected into a shared vector space, and compared through dot products.
Action vectors can be computed ahead of time when an agent repeatedly uses the same tools, interface elements, or game moves. Scoring still grows with the number of candidates, but each additional comparison is a relatively cheap vector operation.
A three-stage training recipe
CLM uses a bidirectional InfoNCE objective over batches of matched state-action pairs. The loss increases similarity between each correct pair and decreases similarity to other examples in the batch. It applies the same operation in reverse, training states to retrieve actions and actions to retrieve states.
This story is for Pro members
You've reached the end of the free preview. Upgrade to AlphaSignal Pro to read the full article - and everything else behind the paywall.
Comments
