State-space models are a low-complexity alternative to transformers for encoding long sequences and capturing long-term dependencies. We propose LOCOST: an encoder-decoder architecture based on state-space models for conditional text generation with long context inputs. With a computational complexity of 𝒪(L log L), this architecture can handle significantly longer sequences than state-of-the-art models that are based on sparse attention patterns. We evaluate our model on a series of long document abstractive summarization tasks. The model reaches a performance level that is 93-96% comparable to the top-performing sparse transformers of the same size while saving up to 50% memory during training and up to 87% during inference. Additionally, LOCOST effectively handles input texts exceeding 600K tokens at inference time, setting new state-of-the-art results on full-book summarization and opening new perspectives for long input processing.

Fine-tuning is the prevalent paradigm to adapt pre-trained language models to downstream tasks. Lightweight fine-tuning methods, such as prefix-tuning, only tune a small set of parameters which alleviates cost. Such methods were shown to achieve results similar to fine-tuning; however, performance can decrease when the inputs get farther from the training domain. Moreover, latest works questioned the efficiency of recent lightweight fine-tuning techniques depending on the task and the size of the model. In this paper, we propose to evaluate the generalization property of prefix-based methods depending on the size of the pre-trained language model in the multi-domain setting on data-to-text generation. We found that their performance depends heavily on the size of the model.