Advertisement
Each step: expand each beam by K candidates; keep top-K total.
What you're seeing
Beam search vs greedy: more chances to find best sequence; K× memory/compute.
★ KEY TAKEAWAY
Beam search keeps top-K hypotheses at each step. More thorough than greedy; slower; mostly used for translation and structured generation now.
▶ WHAT TO TRY
- Click Expand step to see beams branch out.
- Only top-K survive at each level.
- Final answer: best-scoring complete sequence.