Safety
From Dispersion to Attraction: Spectral Dynamics of Hallucination Across Whisper Model Scales
arXiv:2604.08591v1 Announce Type: cross Abstract: Hallucinations in large ASR models present a critical safety risk. In this work, we propose the extit{Spectral Sensitivity Theorem}, which predicts a
arXiv:2604.08591v1 Announce Type: cross Abstract: Hallucinations in large ASR models present a critical safety risk. In this work, we propose the extit{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: intermediate models exhibit extit{Structural Disintegration} (Regime I), characterized by a 13.4% collapse in Cross-Attention rank. Conversely, large models enter a extit{Compression-Seeking Attractor} state (Regime II), where Self-Attention actively compresses rank (-2.34%) and hardens the spectral slope, decoupling the model from acoustic evidence.
Related
- GRM: Utility-Aware Jailbreak Attacks on Audio LLMs via Gradient-Ratio Masking
- Large Language Models Generate Harmful Content Using a Distinct, Unified Mechanism
Source: arXiv cs.AI | 2026-04-13