From Dispersion to Attraction: Spectral Dynamics of Hallucination Across Whisper Model Scales
arXiv:2604.08591v2 Announce Type: replace 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: intermediate models exhibit \textit{Structural
This research provides a theoretical framework 'Spectral Sensitivity Theorem' for understanding and mitigating hallucinations in large AI models, a critical safety and reliability concern as AI adoption accelerates.
Understanding the spectral dynamics of large AI models, particularly the transition from signal decay to rank-1 collapse, is crucial for developing robust and trustworthy AI systems, impacting their commercial viability and deployment.
The theoretical prediction of a phase transition in deep networks offers a new lens for debugging and designing AI models less prone to hallucinations, potentially leading to more reliable AI agents.
- · AI safety researchers
- · AI developers
- · Enterprises adopting AI
- · Companies relying on unreliable 'black box' AI
- · AI model architectures prone to hallucination
Improved methods for detecting and reducing AI hallucination will emerge, increasing trust in AI systems.
More reliable AI models will accelerate the deployment of high-stakes AI applications in industries like healthcare and finance.
The reduced risk of AI hallucinations could diminish regulatory pressure, fostering further innovation and broader AI integration into critical infrastructure.
This signal links to a primary source. Continuum Brief monitors and indexes it as part of the live intelligence stream — we do not republish source content.
Read at arXiv cs.LG