When More is More? When For an LLM is Enough?

In-context length is the LLM’s secret weapon, but with long-context is all changing

Photo by Angely Acevedo on Unsplash

It is better to know some of the questions than all of the answers. — James Thurber

In-context learning (ICL) is one of the most fascinating phenomena of large language models (LLMs). Just provide a few examples and the models can understand the task and execute it with surprising accuracy. Moreover, you do not have to alter a parameter because ICL is performed in inference.

All You Need to Know about In-Context Learning

We still do not really know why it emerges during the training of LLMs but it is the key to the success of LLMs. With the emergence of the long context model, some researchers are beginning to think that it may be the alternative to fine-tuning.

In other words, why not provide a large number of examples and let the model figure out what it needs to do?

RAG is Dead, Long Live RAG

Although this is an attractive alternative we have no idea if it works. After all, the ICL study so far has been conducted only on models with a small context length (most of the models studied had no more than 4K context length). Today’s models have at least 8K, there are open-source models with 32K and closed-source models that claim to have even 1M context length. Despite the publicity proclamations, several researchers question whether the models are really capable of making the best use of this extended context length. Therefore, to be sure that we can efficiently use ICL with a long-context model we should test it