Research map / Large language models / Language, safety and interpretability
$p\textrm{-less}$ Sampling: A Robust Hyperparameter-Free Approach for LLM Decoding
Runyan Tan, Shuang Wu, Phillip Howard
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use p-less instead of top-p | Llama2-7b · GSM8K | AUC: 0.210 -> 0.267 | Improved |
| use p-less instead of min-p | Mistral-7b · GSM8K | AUC: 0.523 -> 0.562 | Improved |
| use p-less instead of min-p | Llama3-70b · CSQA, QASC | AUC: CSQA 0.820 vs 0.819 | No clear effect |
| use p-less instead of min-p | Llama2-7b · GPQA | AUC: 0.243 vs 0.242 | No clear effect |
| use p-less instead of other truncation samplers at temperature 2.0 | Llama-2-7b · Writing Prompts | length-controlled win rate: 48.94 (min-p) -> 65.64 | Improved |
| use p-less instead of top-p, epsilon, eta at temperature 1.5-2.0 | Llama-2-7b · Writing Prompts | length-controlled win rate: top-p 0.00 -> 65.64 at temp 2.0 | Improved |
| use p-less at temperature 1.0 instead of epsilon-sampling | Llama-2-7b · Writing Prompts | length-controlled win rate: 62.18 -> 55.08 | Hurt |
| use p-less instead of min-p for sampling time | Mistral-7b · GSM8K and GPQA | sampling time per token: 0.02497 -> 0.01942 s | Improved |
| use p-less instead of other samplers for generation length | Llama2-7b · CSQA, QASC, GSM8K | generation length | Improved |
| use p-less instead of top-p at temperature 2.0 | Mistral-7b · QASC | diversity: 1.00 -> 0.64 | Improved |
| use p-less_norm variant instead of p-less (ablation) | Llama2-7b, Mistral-7b · math and reasoning | AUC: 0.242 -> 0.248 (GPQA Llama2) | No clear effect |
| p-less vs human annotators default sampling at temperature 2.0 | Llama2-7b · Writing Prompts (human eval) | win rate: 58.8% majority vote | Improved |
Closest papers by meaning
- Turning Up the Heat: Min-p Sampling for Creative and Coherent LLM Outputs (ICLR 2025 · similarity 0.91)
- Hot or Cold? Adaptive Temperature Sampling for Code Generation with Large Language Models (AAAI 2024 · similarity 0.90)
- Sample Smart, Not Hard: Correctness-First Decoding for Better Reasoning in LLMs (ICLR 2026 · similarity 0.90)
- Multi-Branch Self-Drafting for LLM Inference Acceleration (AAAI 2025 · similarity 0.89)
- Judge Decoding: Faster Speculative Sampling Requires Going Beyond Model Alignment (ICLR 2025 · similarity 0.89)
- A Theoretical Perspective for Speculative Decoding Algorithm (NeurIPS 2024 · similarity 0.89)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
