Revisiting Dynamic Evaluation: Online Adaptation for Large Language Models

TL;DR

This paper explores online fine-tuning of large language models during testing, highlighting its role as a form of memory extension that adapts to distributional shifts and blurs the line between in-context learning and fine-tuning.

Contribution

It provides a detailed analysis of online adaptation, emphasizing its connection to memory, and investigates its efficiency, sensitivity, and computational costs in real-world scenarios.

Findings

Online adaptation enhances model performance under distributional shifts.

Speed of adaptation varies with data and model complexity.

Online adaptation can serve as a form of memory, similar to neuroscience concepts.

Abstract

We consider the problem of online fine tuning the parameters of a language model at test time, also known as dynamic evaluation. While it is generally known that this approach improves the overall predictive performance, especially when considering distributional shift between training and evaluation data, we here emphasize the perspective that online adaptation turns parameters into temporally changing states and provides a form of context-length extension with memory in weights, more in line with the concept of memory in neuroscience. We pay particular attention to the speed of adaptation (in terms of sample efficiency),sensitivity to the overall distributional drift, and the computational overhead for performing gradient computations and parameter updates. Our empirical study provides insights on when online adaptation is particularly interesting. We highlight that with online adaptation the conceptual distinction between in-context learning and fine tuning blurs: both are methods to condition the model on previously observed tokens.