Memorizing Transformers
This addresses the issue for AI researchers and practitioners by allowing language models to dynamically incorporate new information without weight updates, though it is incremental as it builds on existing transformer architectures.
The paper tackles the problem of language models needing retraining to acquire new knowledge by enabling them to memorize and use past inputs at inference time, showing that a kNN lookup into a memory of up to 262K tokens improves performance across benchmarks like C4, arXiv, PG-19, Github, and Isabelle.
Language models typically need to be trained or finetuned in order to acquire new knowledge, which involves updating their weights. We instead envision language models that can simply read and memorize new data at inference time, thus acquiring new knowledge immediately. In this work, we extend language models with the ability to memorize the internal representations of past inputs. We demonstrate that an approximate kNN lookup into a non-differentiable memory of recent (key, value) pairs improves language modeling across various benchmarks and tasks, including generic webtext (C4), math papers (arXiv), books (PG-19), code (Github), as well as formal theorems (Isabelle). We show that the performance steadily improves when we increase the size of memory up to 262K tokens. On benchmarks including code and mathematics, we find that the model is capable of making use of newly defined functions and theorems during test time.