Atlas · GenAI 2026
LLM Decoding Strategies
Decoding strategies (temperature, top-p/k, min-p)
conceptPeak: 2023DecodingAI consensus: 2/3
Prerequisites
- mediumTransformer Architecture
Decoding parameters control the sampling from the Transformer's output distribution — understanding the model helps tune its outputs
Recommended reference
Holtzman et al. (2020) 'The Curious Case of Neural Text Degeneration' — ICLR; foundational paper on nucleus sampling (top-p)
Notes from AI deep research
Anthropic Opus
Holtzman (2020) nucleus sampling. Temperature, top-p/k, min-p. Kontrola losowosci dla production
OpenAI Deep Research
Sterowanie losowością i długością [OA#13]
Google Deep Think
Temperature, Top-P/K, Min-P, Repetition Penalty [G#39]
Related skills
- → is part of: LLM Inference Serving(3/3)
- → is part of: Large Language Models (LLM)(3/3)
- → is subcategory of: Prompt Engineering(1/3)