This Google Model Combines Reasoning and Acting in a Single Language Model
ReAct provides an architecture that triggers actions based on language reasoning paths.
I recently started an AI-focused educational newsletter, that already has over 150,000 subscribers. TheSequence is a no-BS (meaning no hype, no news etc) ML-oriented newsletter that takes 5 minutes to read. The goal is to keep you up to date with machine learning projects, research papers and concepts. Please give it a try by subscribing below:
The best source to stay up-to-date with the developments in the machine learning, artificial intelligence, and data…
Reasoning has been one of the emerging capabilities we have seen in the new generation of language pretrained models. Neural network architectures such as GPT-3 or PaLM have excelled in reasoning tasks such as question-answering or even solving mathematical problems. However, most language model still failed to translate reasoning into direct action in a given environment. On the other hand, we have seen models in areas such as embodied tasks or gaming that are incredible efficient at interacting with an environment. Can we combine the two? Recently, Google Research published “ReAct: Synergizing Reasoning and Acting in Language Models” proposing a method that combines reasoning and acting to solve complex language tasks.
The core idea behind ReAct is combine text reasoning and actions in a single model. Reasoning traces are likely to impact the state of a model leading to specific actions while those actions are likely to impact the environment leading to feedback for the reasoning tasks.
As its core architecture, ReAct leverages PaLM-540B as the main language model for both reasoning and action. PaLM is widely considered one of the most complete pretrained language models ever created. In the ReAct architecture, PaLM is prompted with few-shot examples that lead to both…