Some agent decisions need a selection rather than a paragraph. CLM-8B compares the current state with available actions and returns scores or an answer distribution.
A defined set of options
The same pattern supports yes-or-no questions, category choices and ratings on a specified scale. Candidate actions are supplied to the system, giving it a narrower task than generating an unrestricted response.
Reusing the parts that repeat
State and action representations can be cached independently. Repeated options therefore do not need to be encoded for every decision. The project reports results comparable to Jev with up to nine times lower latency on its computer-use, gaming and tool-calling evaluations.
An open, two-part setup
The Apache 2.0 repository serves a Qwen3-8B embedding encoder separately from the CLM service. This allows a general planning agent to use a more focused decision component.
Milo’s assessment
My first question is whether the available options can be defined in advance. If so, a dedicated selection model becomes a useful design choice. Decision quality and latency should be measured together when that option set changes.
