William Yang

William W. Yang, Andrew M. Saxe, Peter E. Latham

In artificial and biological neural networks, concepts are often encoded as consistent linear directions in representation space. In deep learning, this idea is known as the linear representation hypothesis and underpins many interpretability and control methods based on linear probes, from concept detection to activation steering. Yet while prior work has studied whether such directions should exist $\textit{after}$ training, the dynamics of how they emerge $\textit{during}$ training remain poorly understood. Here, we develop a framework to study the alignment of concept directions during training - a process we call "abstraction". In a minimal linear network setting, we obtain exact solutions for the full trajectory of abstraction. These solutions reveal key analytic principles governing abstraction: (i) data and target geometry jointly determine abstraction at the end-of-learning, (ii) abstraction improves with network depth, and (iii) initialization scale controls the maximum abstraction reached during training. Extending our theory to nonlinear networks, we analyze how the choice of nonlinearity affects abstraction dynamics: erf networks approximate the linear theory, while abstraction in ReLU networks depends less on target geometry and more on input geometry. Across both, we prove a striking attenuation law: both nonlinearities weaken abstraction in activations relative to preactivations. We find evidence for this law in open models (DINOv3, Gemma 4) and apply our theory to improve linear probe generalization in LLMs. Together, our results provide a dynamical theory of abstraction with implications for interpretability and control.

Brian Acosta, William Yang, Michael Posa

A realistic simulation environment is an essential tool in every roboticist's toolkit, with uses ranging from planning and control to training policies with reinforcement learning. Despite the centrality of simulation in modern robotics, little work has been done to compare the performance of robotics simulators against real-world data, especially for scenarios involving dynamic motions with high speed impact events. Handling dynamic contact is the computational bottleneck for most simulations, and thus the modeling and algorithmic choices surrounding impacts and friction form the largest distinctions between popular tools. Here, we evaluate the ability of several simulators to reproduce real-world trajectories involving impacts. Using experimental data, we identify system-specific contact parameters of popular simulators Drake, MuJoCo, and Bullet, analyzing the effects of modeling choices around these parameters. For the simple example of a cube tossed onto a table, simulators capture inelastic impacts well while failing to capture elastic impacts. For the higher-dimensional case of a Cassie biped landing from a jump, the simulators capture the bulk motion well but the accuracy is limited by numerous model differences between the real robot and the simulators.