MIT researchers have developed a new method to help. Artificial Intelligence (AI) The systems perform complex reasoning tasks in three areas including coding, strategic planning and robotics.
Major language models (LLMs), including Chat GPT And Claude 3 opus, process and generate text based on human input, called “prompts”. These technologies have improved greatly over the past 18 months, but are limited by their inability to understand context as well as humans or perform well in reasoning tasks, the researchers said.
But MIT scientists now claim to have solved the problem by creating a “treasure trove” of natural language “analyses” that could lead to more powerful AI models. Abstractions turn complex subjects into high-level features and discard trivial information—which can help chatbots reason, learn, understand, and represent knowledge like humans.
Currently, scientists say that LLMs have difficulty summarizing information like a human. However, they have organized natural language abstractions into three libraries in the hopes that they will become more context-aware and respond more human-like.
The scientists detailed their findings in three papers published on the arXiv preprint server on October 30, 2023, December 13, 2023, and February 28, 2023. The first library, calledLibrary involvement from language observations” (LILO) synthesizes, compresses, and documents computer code. The second, named “Acquiring the action domain” (Ada) AI Covers Sequential Decision Making. The Ultimate Framework, Dub”Language Guided Summary” (LGA), helps robots better understand the environment and plan their movements.
Related: 'It would be within its natural right to harm us in self-preservation': how humans can mistreat AI without knowing it at the time.
These papers explore how language can give important context to AI systems so they can handle more complex tasks. They were presented on May 11 at the International Conference on Learning Representations in Vienna, Austria.
“Library learning represents one of the most exciting frontiers in artificial intelligence, offering a way to explore and reason about structured abstractions,” said Robert HawkinsAssistant Professor of Psychiatry at the University of Wisconsin-Madison Statement. Hawkins, who was not involved in the research, added that similar efforts in the past were too computationally expensive to be used on a large scale.
Three of the library's frameworks use neurosymbolic methods — an AI architecture that combines neural networks, which are collections of machine learning algorithms configured to mimic the structure of the human brain — to work, the scientists said. , with a classical program-like logical approach.
Intelligent AI-powered coding
LLMs have emerged as powerful tools for human software engineers, including the likes of GitHub Copilot, but they cannot be used to build full-scale software libraries, the scientists said. To do this, they must be able to organize and compile the code into small programs that are easy to read and reuse, which is where LILO comes in.
The scientists combined a previously developed algorithm that can detect abstractions—known as “Stitch”—with LLMs to create the LILO neurosymbolic framework. Under this system, when an LLM writes code, it is linked to Stich to find abstractions within the library.
Because LILO can understand natural language, it can detect and extract vowels from strings of code and create snowflakes — just like a human software engineer using their common sense. can do.. By better understanding the words used in notation, LLMs could one day create 2D graphics, answer visualization questions, manipulate Excel documents, and more.
Using AI to plan and strategize
LLMs currently can't use reasoning skills to make flexible plans — like the steps involved in cooking breakfast, the researchers said. But the Ada framework, named after the English mathematician Ada Lovelace, may be a way to let these types of assignments adapt and plan when given in a virtual environment.
The framework provided libraries of cooking and gaming projects using LLM to search abstracts from natural language datasets related to these tasks—with the best scored, filtered, and indexed into the library by a human operator. were added. By combining OpenAI's GPT-4 with the framework, scientists overcame the 'code as policies' AI decision-making baseline in performing kitchen simulation and gaming tasks.
By searching for hidden natural language information, the model understood tasks such as putting cold wine in the kitchen cupboard and making the bed—59% and 89%, respectively, compared to performing the same tasks without Ada's influence. with an improvement in the accuracy of . The researchers hope to find other home uses for Ada in the near future.
Giving robots a leg up with the help of AI
The LGA framework also allows robots to better understand their environment in a human-like way – removing unnecessary details from their surroundings and finding better abstractions so they can perform tasks more efficiently.
LGA finds task abstractions in natural language cues such as “bring my hat” with roots based on training footage.
Researchers demonstrated the effectiveness of LGA by using Spot, Boston Dynamics' canine-like quadrilateral robot, to fetch fruit and recycle beverages. Experiments show that robots can effectively scan the world and develop plans in chaotic environments.
Researchers believe that neurosymbolic frameworks like LILO, Ada and LGA will pave the way for “more human-like” AI models by giving them problem-solving skills and allowing them to better navigate their environments.