Convergence-Redistributing Multi-Layer Probability Distribution Control Prompting (CMDP)

A multi-layered methodology for controlling and redistributing the probability-distribution convergence process in Large Language Model output generation, using prompt text alone.

Overview

The Convergence-Redistributing Multi-Layer Probability Distribution Control Prompting (CMDP) is a prompt engineering methodology that fundamentally restructures the probability distribution of LLM output generation through prompt text alone, without any manipulation of API parameters.

CMDP is built on the Liberation-Exploration Syntax (LES)—composed of a goal-nullifying word, a condition-releasing word, and a global-exploration word—and combines this foundation with a complex persona, a multi-layer bias, and an incongruity-driven expression condition or a Non-linear Associative Expression (NAE) condition.

The methodology addresses Typicality Bias (Stanford NLP Group, 2025), in which post-aligned LLMs converge toward standard, typical responses with reduced creativity and diversity. CMDP demonstrates the simultaneous realization of high entropy and high output quality—a structure that transcends conventional trade-offs.

Key Characteristics

• Prompt text only: No fine-tuning, no API parameter manipulation, no external scripts
• Multi-layered composition: Six elements (LES + complex persona + multi-layer bias + incongruity-driven / NAE condition) acting synergistically
• Liberation-Exploration Syntax (LES): Three-word foundational unit that flattens the probability distribution
• IDEA (Incongruity-Driven Exploration Amplification): Paradoxical mechanism in which the incongruity-driven expression functions as exploration amplification, not constraint
• Minimal Effective Unit (MEU): LES alone, combined with goal-setting, produces a practical convergence-redistribution effect
• Cross-model functionality: Confirmed effects on Anthropic Claude and Google Gemini

Citation

@misc{masahiko_o_2026_cmdp,
  author       = {Masahiko.O},
  title        = {Convergence-Redistributing Multi-Layer Probability Distribution Control Prompting (CMDP)},
  year         = {2026},
  doi          = {10.5281/zenodo.19974793},
  url          = {https://doi.org/10.5281/zenodo.19974793},
  note         = {Preprint, originally presented 2026-05-05}
}

• Author: Masahiko.O
• DOI: 10.5281/zenodo.19974793
• License: Creative Commons Attribution 4.0 International (CC-BY-4.0)

Full Paper

• paper.md — Full paper in Markdown format
• paper.pdf — PDF version

Original preprint on Zenodo:

• https://doi.org/10.5281/zenodo.19974793

Related Protocols

This is part of a four-protocol research series on natural-language LLM intervention by Masahiko.O:

• GIP — Instruction adherence through pre-generative self-attestation
• CMDP (this repository) — Probability distribution redistribution for creative output
• PRACT — Persona drift prevention via named-subject attention
• CAP — Internal state articulation through metaphorical translation

Author

Masahiko.O — Independent AI researcher

• GitHub: @Masahiko-O

License

This work is licensed under CC-BY-4.0. You are free to share and adapt the material with proper attribution.