p-less Sampling: A Robust Hyperparameter-Free Approach for LLM Decoding

ICLROral2026

Authors
Runyan Tan, Shuang Wu, Phillip Howard
Affiliation
AI Labs, Thoughtworks
Venue
ICLR 2026
Track
Oral

TL;DR

P-less Sampling: A parameterless sampling strategy grounded in information theory, where the truncation threshold adapts to the entire token probability distribution, is bounded and valid, and dynamically adjusts with temperature.

Opening excerpt from the authors’ abstract. source

Read the paper

Topics

theory llm

← All ICLR 2026 Oral papers · Browse the whole archive