Edge 373: Computationally Efficient LLM Reasoning with ReWOO
In this Issue:
An overview of ReWOO as an LLM reasoning method.
A review of ReWOO’s research paper.
An introduction to LLMFlows, a framework for building LLM applications.
💡 ML Concept of the Day: LLM Reasoning with ReWOO
Augmenting LLMs with external data with techniques such as RAG is one of the most common patterns in generative AI applications. Often overlooked is the impact that RAG-type techniques have on the reasoning abilities of LLMs. Typically, a RAG system needs to coordinate reasoning and action with LLMs. The most common setting is that the LLM reason when to use an external tool. It then pauses to gather the response from the tool before determining the next step. This approach, while simple and straightforward, can sometimes lead to increased computational demands due to repetitive prompts and actions.
ReWOO is a new reasoning technique optimized for information augmented LLMs. ReWOO works by decomposing the fundamental aspects of LLM reasoning and action: step-wise reasoning, tool-calls, and summarization into separate modules. ReWOO achieves this by structing its method around three core components: