How Validation Contracts Change the Label Cost of Reward-Model Replacements
Year: 2026
Type: preprint
Abstract: Replacing a reward model changes which responses a best-of-N system chooses. This paper measures how many labels an audit needs to check that change. It compares guarantees that cover every possible completion of unseen labels with 95% statistical confidence bands. The analysis uses 390 reward-model comparisons across five PPE domains and 630 comparisons on seven Weaver candidate populations, with the source limitations documented in the paper. On PPE, expanding a fresh deterministic audit from the budgets {2, 4, 8, 16, 32} to every integer budget between 2 and 32 adds a paired median of two labels, at full interval width 0.02. Keeping an earlier minimum-cost N=2 audit is more expensive: even the best choice of that starting audit adds a paired median of 300.5 labels compared with a fresh design. An exploratory statistical design reduces expected distinct-label queries by a paired median of 43.66% relative to deterministic validation on PPE. This result describes the tested procedure; it does not establish the cheapest possible statistical audit. All outcomes came from existing datasets. There was no new human annotation or reinforcement-learning training. The reported costs count labels revealed on these fixed datasets. They do not measure annotation labor or dollar savings. The downloads include the paper, editable source and reproducibility files. Extract all four Reward_Model_Replacement_Audit_V1 ZIP files into the same directory and follow README.md. Source-specific rights are listed in SOURCE_RIGHTS.md.
Author: Ozdem, Musa Emin
Language: English
Subfield: Artificial Intelligence
Field: Computer Science
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W7236394302
