Characterize allocation gain as a function of out-of-distribution workload share

Characterize how allocation gain depends on the out-of-distribution share of a workload, including determining the location and explaining the cause of the apparent interior peak observed in the exploratory composition sweep.

Background

The exploratory sweep varies the fraction of clustered, out-of-distribution TSP-100 instances while holding the workload size and budget ratio fixed. For the AM solver, both the signal-free and budget-charged allocation gains increase away from a fully in-distribution workload and decline again at a 50\% out-of-distribution share, suggesting a non-monotone relationship and an interior peak. Because the sweep is descriptive and lacks confirmatory testing, the paper leaves both the peak's location and its mechanism unresolved.

References

We state this and no more; the sweep is descriptive and carries no test, so we do not estimate the location of the peak or offer a mechanism for it.

— Sampling Luck Masquerades as Allocation Gain: Auditing Test-Time Budget Allocation for Neural Combinatorial Optimization  (2608.13087 - Bae, 13 Aug 2026) in Section IV-G, 'Gain against out-of-distribution share'; Section VI, Limitations, item 2