Large language models such as GPT and Llama are trained with a next-token prediction loss. In this work, we suggest that training language models to predict multiple future tokens at once results in higher sample efficiency. More specifically, at each position in the training corpus, we ask the model to predict the following n tokens using n independent output heads, operating on top of a shared model trunk. Considering multi-token prediction as an auxiliary training task, we measure improved downstream capabilities with no overhead in training time for both code and natural language models. The method is increasingly useful for larger model sizes, and keeps its appeal when training for multiple epochs. Gains are especially pronounced on generative benchmarks like coding, where our models consistently outperform strong baselines by several percentage points. Our 13B parameter models solves 12 % more problems on HumanEval and 17 % more on MBPP than comparable next-token models. Experiments on small algorithmic tasks demonstrate that multi-token prediction is favorable for the development of induction heads and algorithmic reasoning capabilities. As an additional benefit, models trained with 4-token prediction are up to 3 times faster at inference, even with large batch sizes.

We study the problem of learning classifiers that perform well across (known or unknown) groups of data. After observing that common worst-group-accuracy datasets suffer from substantial imbalances, we set out to compare state-of-the-art methods to simple balancing of classes and groups by either subsampling or reweighting data. Our results show that these data balancing baselines achieve state-of-the-art-accuracy, while being faster to train and requiring no additional hyper-parameters. In addition, we highlight that access to group information is most critical for model selection purposes, and not so much during training. All in all, our findings beg closer examination of benchmarks and methods for research in worst-group-accuracy optimization.

Attacking Neural Machine Translation models is an inherently combinatorial task on discrete sequences, solved with approximate heuristics. Most methods use the gradient to attack the model on each sample independently. Instead of mechanically applying the gradient, could we learn to produce meaningful adversarial attacks ? In contrast to existing approaches, we learn to attack a model by training an adversarial generator based on a language model. We propose the Masked Adversarial Generation (MAG) model, that learns to perturb the translation model throughout the training process. The experiments show that it improves the robustness of machine translation models, while being faster than competing methods.