Pragmatic Alignment of Large Language Models through Intensity-Weighted Feedback Markers

Large language models (LLMs) exhibit systematic performance degradation in multi-turn interactions, particularly when engaging with expert users whose communicative patterns deviate from the "average user" distribution inherent in RLHF training data. This paper presents an empirical field study documenting a pragmatic alignment methodology developed through extensive in-context learning sessions with Google Gemini. We introduce Pragmatic Intensity Markers (PIMs) (PIMs) - linguistic signals traditionally categorized as vulgar or non-standard - as high-fidelity feedback mechanisms for real-time error correction in LLM interactions. Through systematic observation across extended dialogue sessions (60+ turns), we document a persistent ~60% performance degradation attributable to intent mismatch, KV cache saturation, and regression to prior distributions. Keywords: Large Language Models, Pragmatic Alignment, Intent Mismatch, In-Context Learning, RLHF Limitations, Independent Research