Yan Meng — PhD researcher in NLP at the University of Amsterdam. Multilingual knowledge transfer, reasoning, and translation for language models.https://yanmeng-nlp.com/https://yanmeng-nlp.com/blog/how-does-decoding-affect-reasoning-in-llms/https://yanmeng-nlp.com/blog/how-does-decoding-affect-reasoning-in-llms/Greedy decoding is accurate but repetitive; sampling is diverse but risky. Selective sampling switches between the two at each decoding position to get both.Fri, 11 Jul 2025 00:00:00 GMT