EasyDeepLearn

Why don't chat LLMs use beam search?

medium

Answer

  • Beam search maximizes joint probability but produces 'safe, boring' outputs — highest-probability sequences are generic and repetitive.
  • Human raters strongly prefer sampled outputs (with T + top-p) in open-ended generation.
  • Beam search still shines in constrained-output tasks (structured NER, MT) where the objective is exact-match precision.
  • RLHF and DPO further shift the optimal decoding distribution away from mode-seeking beam search.
Check yourself — multiple choice
  • Beam is standard for chat
  • Beam produces boring, repetitive high-prob outputs; open-ended chat wants sampling with T + top-p
  • Beam is faster
  • Beam is more expressive

Beam: safe/boring outputs; sampling wins for open-ended chat.

#generation#decoding

Practise LLMs & GenAI

214 interview questions in this topic.

Related questions