How Reusable Are Benchmarks with Richer Feedback?
Abstract
We study whether benchmarks reliably guide model selection as developers adapt to evaluation feedback across multiple criteria. We find that the worst-case test-set size needed to estimate the best score among k adaptively chosen models, under any convex combination of the criteria, grows exponentially with the number of criteria, reaching the Θ(√k) cost of answering k adaptive statistical queries with only O(log k) criteria, at fixed accuracy and confidence. In attacks on multi-task large language model benchmarks with five to ten criteria, feedback restricted to nondominated task profiles produces large reused-to-held-out score gaps and frequent false winners. These results challenge a prominent explanation for prior observed reliable benchmark reuse---that developers mainly respond to convincing improvements over the current best---in rich-feedback settings, while leaving open how often ordinary model development encounters this vulnerability.
Related
- From Walks to Skeletons: A Reusable Transport Architecture for Interactive Monte Carlo PDE Solvers
- Condensed PIPG Sequential Convex Optimization for Reusable-Rocket Powered Landing with Strong Aerodynamics
- Semi-Amortized Bayesian Inference with Reusable Local Posteriors
- Reusable QUBO-Derived Block-Schur Inverse Factors for Fixed-Operator Finite-Difference PDE Solvers
- Lag: A Mean-Field Optimal Policy for Online Assignment with Reusable Resources
BibTeX
@misc{allouah2026reusable,
title = {{How Reusable Are Benchmarks with Richer Feedback?}},
author = {Youssef Allouah and John Duchi},
howpublished = {arXiv},
year = {2026},
month = sep,
doi = {10.48550/arxiv.2609.32109},
eprint = {2609.32109},
archivePrefix = {arXiv},
url = {https://arxiv.org/abs/2609.32109},
}
RIS
TY - UNPB
TI - How Reusable Are Benchmarks with Richer Feedback?
AU - Youssef Allouah
AU - John Duchi
PY - 2026
DA - 2026/09/26
PB - arXiv
DO - 10.48550/arxiv.2609.32109
UR - https://arxiv.org/abs/2609.32109
AB - We study whether benchmarks reliably guide model selection as developers adapt to evaluation feedback across multiple criteria. We find that the worst-case test-set size needed to estimate the best score among k adaptively chosen models, under any convex combination of the criteria, grows exponentially with the number of criteria, reaching the Θ(√k) cost of answering k adaptive statistical queries with only O(log k) criteria, at fixed accuracy and confidence. In attacks on multi-task large language model benchmarks with five to ten criteria, feedback restricted to nondominated task profiles produces large reused-to-held-out score gaps and frequent false winners. These results challenge a prominent explanation for prior observed reliable benchmark reuse---that developers mainly respond to convincing improvements over the current best---in rich-feedback settings, while leaving open how often ordinary model development encounters this vulnerability.
ER -