Predicting upcoming events is critical to our ability to effectively interact with our environment and conspecifics. In natural language processing, transformer models, which are trained on next-word prediction, appear to construct a general-purpose representation of language that can support diverse downstream tasks. However, we still lack an understanding of how a predictive objective shapes such representations. Inspired by recent work in vision neuroscience Hénaff et al. (2019), here we test a hypothesis about predictive representations of autoregressive transformer models. In particular, we test whether the neural trajectory of a sequence of words in a sentence becomes progressively more straight as it passes through the layers of the network. The key insight behind this hypothesis is that straighter trajectories should facilitate prediction via linear extrapolation. We quantify straightness using a 1- dimensional curvature metric, and present four findings in support of the trajectory straightening hypothesis: i) In trained models, the curvature progressively decreases from the first to the middle layers of the network. ii) Models that perform better on the next-word prediction objective, including larger models and models trained on larger datasets, exhibit greater decreases in curvature, suggesting that this improved ability to straighten sentence neural trajectories may be the underlying driver of better language modeling performance. iii) Given the same linguistic context, the sequences that are generated by the model have lower curvature than the ground truth (the actual continuations observed in a language corpus), suggesting that the model favors straighter trajectories for making predictions. iv) A consistent relationship holds between the average curvature and the average surprisal of sentences in the middle layers of models, such that sentences with straighter neural trajectories also have lower surprisal. Importantly, untrained models don’t exhibit these behaviors. In tandem, these results support the trajectory straightening hypothesis and provide a possible mechanism for how the geometry of the internal representations of autoregressive models supports next word prediction.

Motor adaptation paradigms provide a quantitative method to study short-term modification of motor commands. Despite the growing understanding of the role motion states (e.g., velocity) play in this form of motor learning, there is little information on the relative stability of memories based on these movement characteristics, especially in comparison to the initial adaptation. Here, we trained subjects to make reaching movements perturbed by force patterns dependent upon either limb position or velocity. Following training, subjects were exposed to a series of error-clamp trials to measure the temporal characteristics of the feedforward motor output during the decay of learning. The compensatory force patterns were largely based on the perturbation kinematic (e.g., velocity), but also showed a small contribution from the other motion kinematic (e.g., position). However, the velocity contribution in response to the position-based perturbation decayed at a slower rate than the position contribution to velocity-based training, suggesting a difference in stability. Next, we modified a previous model of motor adaptation to reflect this difference and simulated the behavior for different learning goals. We were interested in the stability of learning when the perturbations were based on different combinations of limb position or velocity that subsequently resulted in biased amounts of motion-based learning. We trained additional subjects on these combined motion-state perturbations and confirmed the predictions of the model. Specifically, we show that (1) there is a significant separation between the observed gain-space trajectories for the learning and decay of adaptation and (2) for combined motion-state perturbations, the gain associated to changes in limb position decayed at a faster rate than the velocity-dependent gain, even when the position-dependent gain at the end of training was significantly greater. Collectively, these results suggest that the state-dependent adaptation associated with movement velocity is relatively more stable than that based on position.