Frequency Contagion in CDRA
An empirical study of how emotional state propagates through LLM-generated responses. The paper shows that CDRA's constraint layer prevents the model from amplifying the user's emotional frequency while still allowing it to acknowledge and contain it.
Key findings
- Untrained LLMs frequently mirror and amplify user emotional charge.
- CDRA's state classification + behavioral routing blocks the amplification path without flattening tone.
- Frequency contagion is a measurable, reproducible failure mode of default dialogue systems.
Files
Frequency_Contagion_EN_20260629.pdf— Formal paperFrequency_Contagion_EN_20260629.md— Markdown source
Citation
Apert (Jin/Daoqi) and Xiao Han. Frequency Contagion in CDRA: How LLMs Catch and Amplify User State. Zenodo, 2026.
License: CC-BY 4.0
Related work
- CDRA: 10.5281/zenodo.20993162
- Who Am I: 10.5281/zenodo.21056798
Inference Providers NEW
This model isn't deployed by any Inference Provider. 🙋 Ask for provider support