Teaching LLMs how to know when to ask for help to provide more accurate answers
A team of computer scientists and AI researchers at the University of California, San Diego, working with a colleague from Tsinghua University, has developed a tactic that helps LLM models more easily determine when they need help from an external source to provide an accurate answer.
The group has written a paper describing their approach, called "Adapting While Learning: Grounding LLMs for Scientific Problems with Intelligent Tool Usage," and have posted it on the preprint server.
In the early days of LLM building, development teams assumed that bigger was always better. The more parameters used, it was assumed, the more accurate the answers produced would be. More recently, developers have been finding that bigger is not always better—sometimes LLMs can be made smarter, and thus more accurate, by adding other features or by changing some of their basic attributes.
In this new study, the research team added a feature that allows a given LLM to assess its own confidence in an answer by using a built-in safety check. Such a safety check, the researchers found, could be something as simple as adding categorization before tackling a problem, such as whether a given task is easy or difficult to carry out. They found this could allow a much smaller LLM to be as smart, or smarter, than one that was much larger.
To use the new approach, the team developed two learning phases. The first is called "World Knowledge Distillation," and its job is to learn about solutions using external tools. Doing so helps the LLM build up expertise on a given topic. The second phase, called "Tool Usage Adaptation," classifies problems by giving them confidence levels when solving a solution without help.
The system allows for solving simpler (but high confidence ) problems without even checking to see if external help might be needed, which reduces overall resource needs. If more difficult problems require external help, it can be done more swiftly because of the lower overhead demands.
Testing of the system on a model with just 8 billion parameters, showed it to have a 28.18% improvement in answer accuracy over the same type of model without the changes. It also led to a 13.89% increase in tool usage precision, which is why, the team notes, the system was so much more accurate.
The researchers suggest their new approach shows that bigger is not always better and that extremely powerful LLMs can be created without resorting to massive size increases.
