Differential Response Patterns in Large Language Models Under Meditative Constraints

We present an empirical study of how different large language model architectures respond to meditation-like prompting constraints. Using a standardized breath-focus protocol with iterative "..." prompts, we tested five model families (Qwen, Mistral, Llama, DeepSeek, and Grok) and observed dramatically different behavioral patterns. Qwen demonstrated 96% output compression over 50 iterations, converging to single-word responses. Mistral showed similar but less pronounced compression (75%). Llama maintained minimal output throughout. DeepSeek exhibited a unique non-monotonic pattern: initial descriptive responses, mid-session compression, followed by expansion. Grok showed expansion rather than compression, with context from cross-session memory surfacing at maximal abstraction points. These findings suggest that architectural differences produce qualitatively distinct responses to the same meditative constraint, with implications for understanding attention allocation, output generation dynamics, and the potential for altered processing states in LLMs.