The new algorithm helps to improve LLM cooperation for more reasonable and effective solutions.

Have you ever been asked a question to which you knew only part of the answer? To give a more reasoned answer, the best solution would be to call a friend who is better versed in the topic.

The algorithm helps to improve LLM collaboration for more intelligent and effective solutions

"Co-LLM" uses a large general-purpose language model to start responding to the prompt, with a "switchable variable" interfering with certain words to cause a more accurate response from the expert model. Credit: Alex Shipps/MIT CSAIL

Have you ever been asked a question to which you knew only part of the answer? To give a more reasoned answer, the best solution would be to call a friend who is better versed in the topic.

This collaborative process can also help large language models (LLMs) improve accuracy. However, it was difficult to teach LLM to recognize when they should cooperate with another model to respond. Instead of using complex formulas or large amounts of labeled data to explain where models should work together, researchers from the Computer Science and Artificial Intelligence Laboratory of the Massachusetts Institute of Technology (CSAIL) presented a more organic approach.

Their new algorithm, called "Co-LLM", can combine the basic general-purpose LLM with a more specialized model and help them work together. While the first one creates an answer, Co-LLM reviews each word (or token) in its answer to see where it can call for a more accurate response from the expert model. This process leads to more accurate answers to things such as medical clues, mathematical and logical problems. Since an expert model is not needed at each iteration, it also leads to more effective response generation.

To decide when the base model needs the help of an expert model, the framework uses machine learning to train a "switching variable" or a tool that can indicate the competence of each word in the responses of two LLMs. The switch is similar to a project manager who finds the areas in which a specialist should be called.

If you asked Co-LLM to name several examples of extinct bear species, for example, the two models would have made the answers together. The general purpose LLM begins to make up the answer, and the switch variable interferes in those parts where it can insert the best token from the expert model, for example, add the year when the bear species is extinct.

"With Co-LLM, we essentially train general-purpose LLM to "phone" the expert model when necessary," says Shannon Shen, a MIT graduate student in electrical engineering and computer science and a branch of CSAIL, the lead author of a new article on the approach. The results are published on the arXiv preprint server.

The algorithm helps to improve LLM collaboration for more intelligent and effective solutions

"Co-LLM" uses a large general-purpose language model to start responding to the prompt, with a "switchable variable" interfering with certain words to cause a more accurate response from the expert model. Credit: Alex Shipps/MIT CSAIL

Have you ever been asked a question to which you knew only part of the answer? To give a more reasoned answer, the best solution would be to call a friend who is better versed in the topic.

This collaborative process can also help large language models (LLMs) improve accuracy. However, it was difficult to teach LLM to recognize when they should cooperate with another model to respond. Instead of using complex formulas or large amounts of labeled data to explain where models should work together, researchers from the Computer Science and Artificial Intelligence Laboratory of the Massachusetts Institute of Technology (CSAIL) presented a more organic approach.