Sampling Strategies: Greedy, Multinomial, Beam Search & Contrastive Search
~20 min read
Even with a probability distribution over next tokens, you still need a strategy to actually pick one. The 4 main strategies — greedy, multinomial, beam search, contrastive search — trade off speed, coherence, and diversity very differently.
Sampling Strategies: Greedy, Multinomial, Beam Search & Contrastive Search is a Pro topic
Sign in, then upgrade to Pro or Power to unlock this topic and the full AI Engineering curriculum.
Key points
- •Greedy: always picks the single highest-probability token — fast, deterministic, but prone to repetitive loops
- •Multinomial: samples from the probability distribution (shaped by temperature) instead of always taking the top token
- •Beam search: keeps the top-k PARTIAL sequences alive at each step, approximating whole-sequence maximization instead of greedy per-token choices
- •Beam search is widely used where correctness matters more than creativity, e.g. machine translation
- •Contrastive search: penalizes candidates too similar to what's already generated, balancing fluency with diversity to avoid repetition loops