On Retrieval Augmentation and the Limitations of Language Model Training

arXiv:2311.09615v2 Announce Type: replace
Abstract: Augmenting a language model (LM) with $k$-nearest neighbors ($k$NN) retrieval on its training data alone can decrease its perplexity, though the underlying reasons for this remain elusive. In this work, we rule out one previously posited possibility -- the "softmax bottleneck." We then create a new dataset to evaluate LM generalization ability in the setting where training data contains additional information that is not causally relevant. This task is challenging even for GPT-3.5 Turbo. We …

abstract arxiv augmentation cs.cl data language language model language model training limitations neighbors perplexity possibility retrieval softmax training training data type work