Explore beam-search decoding for Pairton

Investigate whether more sophisticated sampling algorithms, such as beam search, can improve upon Pairton’s greedy inference procedure for selecting particle-assignment edges.

Background

Pairton approximates the maximum a posteriori particle assignment by greedily selecting edges that maximize a sum of predicted logits at each reconstruction step. This decoding strategy is computationally simple and deterministic, but it is only a heuristic approximation to the globally optimal assignment.

The paper explicitly identifies beam search as a possible alternative sampling strategy and leaves its exploration unresolved. Evaluating such methods could determine whether broader search over competing partial assignments improves reconstruction performance without compromising inference efficiency.