In plain words: Finetune a language model on text from one year or month, then subtract the original weights to get a direction that captures that time period. Blending directions from nearby periods makes a model that handles in-between and later periods better, without the usual retraining.
Abstract · Time is Encoded in the Weights of Finetuned Language Models
We present time vectors, a simple tool to customize language models to new time periods. Time vectors are created by finetuning a language model on data from a single time (e.g., a year or month), and then subtracting the weights of the original pretrained model. This vector specifies a direction in weight space that, as our experiments show, improves performance on text from that time period. Time vectors specialized to adjacent time periods appear to be positioned closer together in a manifold. Using this structure, we interpolate between time vectors to induce new models that perform better on intervening and future time periods, without any additional training. We demonstrate the consistency of our findings across different tasks, domains, model sizes, and time scales. Our results suggest that time is encoded in the weight space of finetuned models.
Kai Nylund, Suchin Gururangan, Noah A. Smith
arXiv:2312.13401 · cs.CL · submitted Dec 20, 2023 · updated Dec 30, 2023
abstract · pdf · html · Added references to Jaidka et al. (2018) and Loureiro et al. (2022)
Feels like a click bait title. Of course language model weights encode different writing styles. The fact that you can lift out a vector to stylize writing is also more interesting, but that’s also nothing newly discovered here. It should be obvious that this is possible given that you can prompt ChatGPT to change its writing style.