Words in Motion: Extracting Interpretable Control Vectors for Motion Transformers

Transformer-based models generate hidden states that are difficult to interpret. In this work, we aim to interpret these hidden states and control them at inference, with a focus on motion forecasting. We use linear probes to measure neural collapse towards interpretable motion features in hidden states. High probing accuracy implies meaningful directions and distances between hidden states of opposing features, which we use to fit interpretable control vectors for activation steering at inference. To optimize our control vectors, we use sparse autoencoders with fully-connected, convolutional, MLPMixer layers and various activation functions. Notably, we show that enforcing sparsity in hidden states leads to a more linear relationship between control vector temperatures and forecasts. Our approach enables mechanistic interpretability and zero-shot generalization to unseen dataset characteristics with negligible computational overhead. Our implementation is available at https://github.com/kit-mrt/future-motion