Your team wants to upgrade from an 8k-context model to...

Simply increasing the context window from 8k to 128k does not automatically resolve a lost-in-the-middle regression. A larger window represents increased capacity, but it does not guarantee that the model has the comprehension capabilities to effectively reason over that much data.

To verify if the upgrade is effective, you must confirm that the new checkpoint was trained to handle that specific length. If the model was not trained for the extended context, it will likely continue to struggle with information retrieval and positional bias, regardless of the increased token limit.