Whisper models show a scale-dependent spectral phase transition from mid-size cross-attention rank collapse under stress to large-model self-attention compression that decouples from acoustic evidence.
Transformers learn low sensitivity functions: Investigations and implications
1 Pith paper cite this work. Polarity classification is still indexing.
1
Pith paper citing it
fields
cs.LG 1years
2026 1verdicts
UNVERDICTED 1representative citing papers
citing papers explorer
-
From Dispersion to Attraction: Spectral Dynamics of Hallucination Across Whisper Model Scales
Whisper models show a scale-dependent spectral phase transition from mid-size cross-attention rank collapse under stress to large-model self-attention compression that decouples from acoustic evidence.