Quantify FLEET’s Parallel Speedup

Quantify the parallel speedup achieved when multiple FLEET workers coordinate through shared search memory and explore distinct high-scoring branches.

Background

FLEET supports coordinated parallelization in which workers can reserve or receive distinct high-scoring branches from shared search memory. The intended benefit is to reduce redundant exploration across concurrently generated trajectories.

The paper describes this capability architecturally but does not measure its performance. Quantifying the resulting speedup would determine the practical scalability of coordinated FLEET workers and establish how effectively shared search-state information translates into reduced wall-clock inference cost.