Abstract
Large language models (LLMs) engaged in extended dialogues may exhibit "tone inertia," where a model's style can become fixed and may not adapt when a conversational topic shifts, potentially leading to contextually inappropriate responses. A system can manage this behavior by algorithmically decoupling conversational context into components, such as a factual context for semantic knowledge and a stylistic context for tone and stylistic profile. A domain shift detection module can identify shifts in the conversation based on mathematical/vector detection, which may trigger a reset of the stylistic context while the factual context is preserved. This process can allow dynamic adaptation of an LLM’s stylistic profile, for example, from technical to empathetic, without manual user intervention or loss of conversational history.
Creative Commons License
This work is licensed under a Creative Commons Attribution 4.0 License.
Recommended Citation
Doshi, Viraj, "Decoupling Stylistic and Factual Context in LLMs to Mitigate Conversational Tone Inertia", Technical Disclosure Commons, (March 30, 2026)
https://www.tdcommons.org/dpubs_series/9657