What question did this study set out to answer?

The research aims to explore the limitations of reasoning models in maintaining control over their reasoning processes.

March 29, 20260 citationsOpen Access

Spectral Language Degradation in Chain-of-Thought: Why Reasoning Models Cannot Control Their Own Reasoning

Key Points

The research aims to explore the limitations of reasoning models in maintaining control over their reasoning processes.
Analysis of reasoning models' chains of thought
Application of spectral language hierarchy
Comparison of phonological and semantic structures
Examination of faithfulness gap between CoT and true reasoning
Chain-of-thought reflects surface phonology but deviates from actual computations
Faithfulness gap identified between phonological and semantic elements
Degradation of CoT fidelity predicted to occur as 1/Pe
At Pe ≥ 4, CoT collapses to a single phoneme, obscuring deep computation

Abstract

OpenAI's finding that reasoning models struggle to control their chains of thought — and that stated reasoning does not always reflect true computation — is predicted by the spectral language hierarchy (HP53, HP59). The eigenvalue spectrum constitutes a four-level language (phonology→syntax→semantics→pragmatics) that degrades bottom-up under Pe. Chain-of-thought is surface phonology (Method B: collapse from 99 to 1 phoneme as Pe increases), while the model's actual computation is deep grammar (Method C: quasi-modular ring). The faithfulness gap between CoT and true reasoning is the spectral gap between phonological and semantic destruction exponents (α=1.31 vs α=1.96). Predictions: CoT faithfulness degrades as 1/Pe; at Pe≥4, CoT becomes a monophone — surface compliance masking arbitrary deep computation.

KI fragen

Bookmark

View Full Paper