Life After Benchmark Saturation: A Case Study of CORE-Bench

2026-06-23

AI Summary

Traditional benchmarks are often abandoned once accuracy saturates, overlooking the need for in-depth multidimensional evaluation of agent construct validity, out-of-distribution generalization, efficiency, reliability, the relative contributions of models versus scaffolding, and human-AI collaboration gains. Taking CORE-Bench Hard as a case study, this work proposes a multidimensional evaluation framework that transcends accuracy by introducing an improved benchmark, CORE-Bench v1.1, and a dedicated out-of-distribution task suite, CORE-Bench OOD, alongside randomized controlled trials to quantify human-AI collaboration effects. The study uncovers construct validity threats in the original benchmark, demonstrates that the revised benchmark effectively captures efficiency and reliability, and reveals approximately 2× acceleration in real-world replication tasks through human-AI collaboration, thereby affirming the continued research value of saturated benchmarks.

Abstract

When a benchmark's accuracy saturates, it is often retired and replaced with a more challenging version. We show that this approach privileges accuracy and misses the opportunity to study six other key dimensions of agent performance: construct validity issues such as shortcuts, out-of-distribution generalizability, efficiency, reliability, the relative importance of the model versus the scaffold, and uplift from human-agent collaboration. We use CORE-Bench Hard, a benchmark for computational reproducibility of scientific code, as a case study to demonstrate that measuring agents along these dimensions yields meaningful insights into agent performance even after accuracy saturates. First, we surface threats to construct validity in CORE-Bench Hard that are difficult to anticipate with less capable agents. We introduce an improved benchmark, CORE-Bench v1.1, and an out-of-distribution task suite, CORE-Bench OOD. Second, we find that despite accuracy saturation, CORE-Bench v1.1 remains useful for measuring efficiency, reliability, model performance, and scaffold performance. Finally, we conduct a small-scale randomized experiment to measure uplift from human-agent collaboration on real-world computational reproducibility tasks. We find a statistically significant speedup by about a factor of two -- likely underestimated due to one-fifth of human-only reproductions reaching the time limit before completing -- and describe various other findings. Together, our contributions present a more rigorous alternative to the dominant accuracy-centric evaluation paradigm.

Problem

• benchmark saturation
• construct validity
• out-of-distribution generalization
• agent evaluation
• human-agent collaboration

Innovation

• construct validity
• out-of-distribution generalization
• human-agent collaboration
• benchmark saturation
• computational reproducibility