Enable javascript in your browser for better experience. Need to know to enable it? Go here.
Research

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

p-less sampling replaces fixed truncation hyperparameters like top-p and min-p with a threshold derived from the full token distribution, grounded in Renyi entropy. Tested on three LLMs across math, reasoning and creative writing tasks, it holds accuracy steady at high temperatures where other methods degrade, while also sampling faster and producing shorter generations.