THINK-AND-EXECUTE is a novel framework that enhances reasoning in LLMs by dividing the task process into two phases: THINK and EXECUTE. In the THINK phase, an Instructor LM generates pseudocode that encapsulates the essential logic needed to solve a task, informed by examples from similar tasks. This pseudocode then guides the subsequent EXECUTE phase, where a Reasoner LM carries out the task using the established logic. This methodology aims to improve the efficiency and accuracy of LLMs in handling diverse tasks by focusing on shared reasoning processes.
The THINK phase involves the Instructor LM generating pseudocode that captures the underlying logic of a task, facilitating step-by-step reasoning for multiple instances.
By utilizing a meta prompt with examples, the Instructor LM analyzes the target task and identifies essential reasoning logic that transcends individual task instances.
In our methodology, we believe that constructing task-level pseudocode through the Instructor LM can significantly enhance the reasoning capabilities of LLMs when executing tasks.
The EXECUTE phase, driven by the Reasoner LM, applies the pseudocode generated in the THINK phase to carry out the specific tasks effectively.
Collection
[
|
...
]