AI researchers have been developing systems that can speak natural language with the same elegance and adaptability as people since the field’s inception. Although very simple models, such as Eliza from 1966, may provide responses to some plausible claims, it has always been relatively easy to ask questions that reveal their shortcomings compared to people – their lack of actual ‘understanding’. Although large language models (LLMs) such as GPT-4 and ChatGPT greatly outperformed expectations a few years ago, they are similar. The internet is full of people who take great pleasure in tinkering with ChatGPT to produce output that a 5-year-old might see as unwise.
This behavior should come as no surprise, given how LLMs are created and taught. They are not designed with understanding in mind. They have been taught to produce word sequences that, given the context, seem believable to humans. According to Mahwald et al., legal scholars have mastered the art of linguistic competence, or knowing how to say things, but they need to be more adept at functional competence or understanding what needs to be said. In particular, they can be (relatively) easily fooled into, for example, asking for an answer to a simple mathematical problem not included in their training suite or asking for a solution to a unique planning problem that requires knowledge of how the outside world works.
Do they now need to work harder to incorporate all math and planning tasks into their training suite? This is a fool’s job. But why should it be necessary, on the other hand? They already have general purpose symbolic charts and calculators that are guaranteed to produce accurate results. Linking LLM to such technologies is a logical alternative strategy that they weren’t the first to look into. With this purpose in mind, the research described in this paper aims to provide LLM with the first ever accurate solution to planning difficulties. They want to do this even with fine tuning without changing the LLMs themselves.
🚀 Join the fastest ML Subreddit community
Instead, researchers from UT Austin and the State University of New York introduced a method known as LLM + P which, when given a natural language description of the planning problem, is LLM:
- An appropriate problem description emerges as input to the general purpose diagram.
- Solve the problem using a general purpose planner.
- Converts schema production back to natural language.
In this work, they do not require the LLM to understand when to make a claim that could be addressed by the proposed LLM+P pipeline. Recognizing when an LLM+P must deal with a claim will be important for future research. Their comprehensive empirical analyzes show that LLM+P can accurately answer many more planning problems than LLMs alone. This broad technique can be used to answer any class of cases for which there is a good and comprehensive solution, such as arithmetic problems (using calculators), although it is demonstrated in this work on planning problems. The code and results are publicly available on GitHub.
scan the paper And github link. Don’t forget to join 20k+ML Sub RedditAnd discord channelAnd Email newsletter, where we share the latest AI research news, cool AI projects, and more. If you have any questions regarding the above article or if we’ve missed anything, feel free to email us at Asif@marktechpost.com
🚀 Check out 100’s AI Tools in the AI Tools Club
Anish Teeku is a Consultant Trainee at MarktechPost. He is currently pursuing his undergraduate studies in Data Science and Artificial Intelligence from the Indian Institute of Technology (IIT), Bhilai. He spends most of his time working on projects aimed at harnessing the power of machine learning. His research interest is in image processing and he is passionate about building solutions around it. Likes to communicate with people and collaborate on interesting projects.