From Dispersion to Attraction: Spectral Dynamics of Hallucination Across Whisper Model Scales

A new theoretical framework proposes that hallucinations in large automatic speech recognition models follow a predictable phase transition, moving from signal decay to a collapsed attractor state as model scale increases, according to research posted to arXiv [1][2]. Kirill Borodin, the author of the study, introduces the Spectral Sensitivity Theorem to describe how deep networks shift between two regimes under adversarial stress [1][2]. The theorem predicts a transition from a dispersive regime, marked by signal decay, to an attractor regime characterized by a rank-1 collapse, governed by layer-wise gain and alignment [2]. The work focuses on OpenAI’s Whisper models, ranging from Tiny to Large-v3-Turbo [1][2]. Analysis of activation graph eigenspectra confirmed the theoretical prediction [2]. Intermediate models exhibited what the paper terms Structural Disintegration, or Regime I, which included a 13.4% collapse in Cross-Attention rank [1][2]. In contrast, large models entered a Compression-Seeking Attractor state, or Regime II, where Self-Attention actively compressed rank by -2.34% and hardened the spectral slope, effectively decoupling the model from acoustic evidence [1][2]. The paper was submitted to arXiv on March 31, 2026, and revised on June 26, 2026 [1]. arXiv, which began operating in 1991, serves as an open-access repository for electronic preprints that are moderated but not peer-reviewed, and it now receives about 24,000 submissions per month [6]. The platform also hosts community-developed tools through its arXivLabs framework, which provides features such as bibliographic explorers and code finders on article pages [4][5]. The study’s abstract page includes links to several of these tools, including Bibliographic Explorer and Connected Papers [1]. The research addresses a critical safety risk posed by hallucinations in large ASR models [1][2]. By identifying the spectral signatures of these two regimes, the theorem offers a potential diagnostic for when a model begins to fabricate outputs rather than transcribe speech. The findings have not yet been published in a peer-reviewed journal [6].