On the Generalization Ability of Retrieval-Enhanced Transformers
This work addresses the evaluation challenges for retrieval-augmented language models, highlighting potential overestimation of generalization in incremental research.
The paper investigates the generalization ability of Retrieval-Enhanced Transformers (RETRO), finding that performance gains largely stem from token overlap between the database and test data, indicating less non-trivial generalization than previously thought.
Recent work on the Retrieval-Enhanced Transformer (RETRO) model has shown that off-loading memory from trainable weights to a retrieval database can significantly improve language modeling and match the performance of non-retrieval models that are an order of magnitude larger in size. It has been suggested that at least some of this performance gain is due to non-trivial generalization based on both model weights and retrieval. In this paper, we try to better understand the relative contributions of these two components. We find that the performance gains from retrieval largely originate from overlapping tokens between the database and the test data, suggesting less non-trivial generalization than previously assumed. More generally, our results point to the challenges of evaluating the generalization of retrieval-augmented language models such as RETRO, as even limited token overlap may significantly decrease test-time loss. We release our code and model at https://github.com/TobiasNorlund/retro