Stanford/NVIDIA researchers unveil CLM, a contrastive-learning model for fast agent action selection
CLM-8B matches larger 'System Two' model Jev's accuracy while running up to 9× faster on computer-use, gaming and tool-calling tasks.
What to know
- CLM pairs a state encoder and an action encoder trained with a contrastive InfoNCE objective, letting the model score and select actions directly via embedding similarity rather than generation.
- CLM-8B reportedly matches the accuracy of the larger Jev model across computer-use, gaming, and tool-calling benchmarks while running up to 9× faster.
- Speed gains are largest in tasks with many candidate actions (e.g., WikiRacing) or high action reuse across states (e.g., a T-Rex game benchmark).
- The work comes from a Stanford/NVIDIA Research team including Jacky Kwok, Christopher Ré, and Azalia Mirhoseini; announced via a Notion writeup and a viral X post.
Jacky Kwok Lead researcher, Stanford UniversityChristopher Ré Senior co-author, Stanford UniversityAzalia Mirhoseini Senior co-author, Stanford University / NVIDIA ResearchHangoo Kang Co-author, Stanford UniversityTarun Suresh Co-authorJon Saad-Falcon Co-author
How it unfolded 3 developments, newest first · click a bar or a number to jump posts
-
3
Coverage of CLM spreads further on Hacker News
A second Hacker News submission linking to the original X announcement appears, indicating continued pickup of the CLM announcement, though with minimal additional discussion.
-
2
Research team publishes CLM details and benchmark results
A Notion page from the Stanford/NVIDIA team (Jacky Kwok, Hangoo Kang, Tarun Suresh, Jon Saad-Falcon, Marco Pavone, Christopher Ré, Azalia Mirhoseini) explains CLM's contrastive state/action encoder design and reports CLM-8B matching Jev's performance at up to 9× the speed across computer-use, gaming, and tool-calling tasks; the writeup circulates on Hacker News.
“Across computer-use, gaming, and tool-calling tasks, CLM-8B performs on par with Jev while running up to 9× faster.”
— Stanford/NVIDIA research team, Paper authors · source -
1
Jacky Kwok announces CLM-8B on X
Lead researcher Jacky Kwok posts an introduction of Contrastive Language Model (CLM), describing it as an 'ultra-fast System One Model' pre-trained on internet-scale data that is up to 9× faster than Jev with comparable performance.
“Introducing Contrastive Language Model (CLM): an ultra-fast System One Model trained with a contrastive learning objective that connects states and actions.”
— @jackyk02 -
first by MarkTechPost, 4h ago
-
Introducing Contrastive Language Model (CLM): an ultra-fast System One Model trained with a contrastive learning objective that connects states and actions. CLM-8B is pre-trained on internet-scale data and delivers up to 9× faster inference than Jev ⚡ while achieving comparable
-