Why Two Models Are Better Than One

You may have heard of the saying, “two heads are better than one.” The same could be said about large language models.

Yes, developers have figured out that model performance can be improved by combining a couple of LLMs. The concept behind "model merging" is surprisingly intuitive. Developers combine the weights (or the “settings” that determine how a model responds to queries) of two or more models trained for different purposes to create a single model that exhibits the strengths of the original models without additional training. For instance, Tokyo-based model developer Sakana AI combined an LLM that generates text in Japanese with a math-focused LLM to create a new LLM that’s especially adept at solving math problems in Japanese. (See more about the Khosla- and Lux-backed Sakana here.)

Until now, model merging has been considered a “black art,” as developers often have to guess how heavily they should weigh one model or the other when combining them, said Sakana AI co-founder David Ha.