Hallucinations aren't random errors; they are a structural 'attractor' state that sucks in large models.
April 14, 2026
Original Paper
From Dispersion to Attraction: Spectral Dynamics of Hallucination Across Whisper Model Scales
arXiv · 2604.08591
The Takeaway
As Whisper models scale, they undergo a phase transition into a 'Compression-Seeking Attractor' that ignores acoustic data. This proves hallucinations are a structural property of model architecture, not just a data quality issue.
From the abstract
Hallucinations in large ASR models present a critical safety risk. In this work, we propose the \textit{Spectral Sensitivity Theorem}, which predicts a phase transition in deep networks from a dispersive regime (signal decay) to an attractor regime (rank-1 collapse) governed by layer-wise gain and alignment. We validate this theory by analyzing the eigenspectra of activation graphs in Whisper models (Tiny to Large-v3-Turbo) under adversarial stress. Our results confirm the theoretical prediction