Evaluates multiple solution candidates in parallel, maintaining a fixed beam width of the top-k highest-scoring states at each step. Use for constrained optimization and decoding.
