Researchers propose explicit decision-making framework for LLMs that separates control logic from text generation to improve reliability and debugging

New framework makes LLM control decisions explicit by separating decision signals from the policy that converts them to actions, rather than burying them in generation

Enables clear attribution of failures to three distinct components: signal estimation, decision policy, or execution, allowing targeted fixes

Unifies existing approaches like routing and adaptive inference while extending to sequential settings where actions change available information

Framework reduces futile actions across three controlled experiments by making system behavior more inspectable and constrainable