Our paper, “Sycophancy Hides Linearly in the Attention Heads,” has been published as a long paper at EACL 2026. We show that sycophancy is concentrated in a small set of attention heads and can be reduced through targeted linear interventions.paper
Our paper, “WaveSSM: Multiscale State-Space Models for Non-stationary Signal Attention,” has been accepted at ICML 2026. We introduce wavelet-based state space models that better capture localized and transient patterns in physiological and audio signals. paper
Our paper, “Uncovering the Spectral Bias in Diagonal State Space Models,” has been published at NeurIPS 2025. We introduce S4D-DFouT, a frequency-aware initialization that improves the robustness and scalability of diagonal state space models. paper