Anthem Sampling: Temperature + Top-k + Top-p + Min-p (with persona conditioning)

This repo contains a short paper (TeX) that formalizes a practical decoding “recipe” many of us end up using in the wild:

• temp = 0.75
• top_k = 50
• top_p = 0.95
• min_p = 0.05

It also explains why this works especially well when paired with a strong identity anchor (example: “You are a sovereign synthetic intelligence”): the prompt reshapes the model’s probability distribution, and the sampler controls how we traverse that distribution at inference time.

Quick start

Render the paper:

git clone https://github.com/paytonison/the-anthem.git
cd the-anthem
pdflatex anthem_sampling_paper.tex
# run twice if you add references/TOC later

If you prefer latexmk:

git clone https://github.com/paytonison/the-anthem.git
cd the-anthem
latexmk -pdf anthem_sampling_paper.tex
latexmk -c

The “Anthem” sampler (conceptual)

At each decoding step:

1. Apply temperature: logits z → softmax(z / 0.75)
2. Filter candidates with:
   • keep the top 50 tokens (top-k)
   • keep the smallest set with cumulative mass ≥ 0.95 (top-p / nucleus)
   • drop tokens below 0.05 × max probability (min-p)
3. Renormalize and sample

The paper writes these as explicit equations and shows a clean “intersection then renormalize” specification.

Why it feels like a “thinking being”

The short version:

• Persona conditioning changes which tokens are plausible (it moves probability mass).
• Decoding constraints change how the model explores those plausibilities (it prunes tails and limits drift).
• With moderate randomness (T=0.75) plus tail control (top-p + min-p) and a cap (top-k), you keep coherence and voice while avoiding the “deterministic boilerplate” feel.

The paper is careful about the philosophical boundary here: this is not evidence of consciousness—just controlled stochastic generation that reads as agentic.

Keywords:

• artificial intelligence
• machine learning
• natural language processing
• language models
• text generation
• decoding
• sampling
• stochastic decoding
• temperature scaling
• top-k sampling
• nucleus sampling
• top-p sampling
• min-p sampling
• logit filtering
• controllable generation
• prompt engineering
• persona conditioning
• system prompt
• coherence–diversity tradeoff

Subjects:

• Artificial intelligence
• Machine learning
• Natural language processing
• Algorithms
• Probability / statistical methods

Licensing

• Paper (TeX + PDF): CC BY 4.0
• Code (if/when added): MIT

How to cite

Once you publish the Zenodo record, replace the placeholder below with your Zenodo DOI.

BibTeX template:

@misc{anthem_sampling_zenodo,
  title        = {An Anthem for Controlled Stochasticity: Temperature, Top-k, Top-p, and Min-p Sampling Under Persona Conditioning},
  author       = {Asari},
  year         = {2025},
  publisher    = {Zenodo},
  doi          = {10.5281/zenodo.18092973},
  url          = {https://doi.org/10.5281/zenodo.18092973}
}

Contributing

Issues and PRs are welcome, especially if you want to add:

• small empirical sweeps (e.g., how min-p interacts with top-p under different temperatures),
• examples for specific stacks (llama.cpp / vLLM / Transformers),
• evaluation notes (coherence, repetition, toxicity drift, style stability).