The Benchmark Ceiling 0530version
The paper argues that AI benchmarks, the primary instruments for measuring and governing AI capability, suffer from a structural "benchmark ceiling" problem. As foundation models improve, they quickly saturate the easy majority of benchmark items, leaving only a thin tail of hard items that retain discriminating power. These hard items require elite, low‑codifiability human judgment, which is scarce and expensive. The authors formalize this dynamic with a model showing (1) benchmark signal depreciation as validity declines, (2) concentration of valid signal in hard‑tail items, (3) convexly rising replacement costs for such items, and (4) private underinvestment relative to the social optimum. Empirical evidence from the micro1 platform demonstrates a right‑skewed wage premium for evaluators who work on high‑difficulty, low‑codifiability tasks, confirming the scarcity of the evaluative 1 %. The paper then analyzes the political‑economy of benchmark production, highlighting public‑goods underinvestment, capture incentives, and the trade‑off between item‑level transparency (which raises contamination risk) and procedural transparency (which improves accountability). Governance recommendations include protected live item pools, transparent procedures, independent benchmark institutions, sustained public or quasi‑public funding, and mandatory disclosure of benchmark design, evaluator credentials, and contamination testing. The work concludes that reliable AI governance depends on well‑funded, independent evaluation infrastructure built on scarce expert judgment, and that current market failures threaten the credibility of AI capability claims.
Topics
Benchmark Ceiling Problem: Saturation of Easy Items and Reliance on Hard‑Tail Human Judgment
Defines the core thesis that as AI models improve, easy benchmark items become uninformative, concentrating discriminating signal in a small set of hard items that require elite expert judgment.
Formal Model of Benchmark Signal Depreciation and Validity
Presents the mathematical framework (Equations 1‑4, Propositions 1‑4) linking benchmark scores to latent model quality, showing how validity Bₜ endogenously determines measurement error and how it declines with model capability and contamination.
Difficulty Distribution and Hard‑Tail Concentration
Describes empirical and theoretical evidence that benchmark items follow a log‑normal difficulty distribution, with the valid signal shifting toward items whose difficulty matches frontier capability.
Characteristics of Hard Benchmark Items and the Evaluative 1 %
Identifies four structural features of hard items (deep domain knowledge, low codifiability, contested ground truth, resistance to pattern‑matching) and argues that only a thin stratum of evaluators can produce them.
Rising Replacement Cost of Frontier‑Discriminating Items
Shows that producing items with difficulty θ≈cₜ incurs convex costs C(θ), implying that as models advance, the cost of maintaining benchmark validity rises sharply.
Public‑Goods Underinvestment and Capture Incentives in Benchmark Production
Analyzes the political economy: benchmarks are non‑rival public goods, private producers underinvest (Proposition 4) and have incentives to shape benchmarks toward their own strengths (Proposition 5).
Empirical Evidence of Scarcity Premium from the micro1 Platform
Presents data on over 1,000 credentialed professionals showing a right‑skewed wage premium for high‑difficulty tasks, confirming the scarcity of elite evaluators and the convex cost assumption.
Governance Trade‑offs: Item‑Level vs Procedural Transparency
Explores how revealing live items increases contamination risk, while procedural transparency (governance, evaluator credentials) improves accountability; proposes protected item pools with transparent procedures.