Revisiting Dynamic Evaluation: Online Adaptation for Large Language Models
This work addresses the problem of adapting language models to distributional shifts during inference for AI researchers, but it is incremental as it revisits and refines existing dynamic evaluation concepts.
The paper investigates online fine-tuning of large language models at test time, known as dynamic evaluation, to improve predictive performance under distributional shift, and finds that it enhances sample efficiency and blurs the distinction between in-context learning and fine-tuning.
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.