Imagine a race where the finish line keeps moving. That's the current state of Large Language Model (LLM) evaluation. New benchmarks pop up constantly, each claiming to measure 'intelligence' or 'reasoning,' but how do we know if they're any good? Researchers often use Benchmark Agreement Testing (BAT) to see if new benchmarks align with established ones. But there's a problem: the very process of testing benchmarks is inconsistent and often leads to misleading results. A new research paper, "Do These LLM Benchmarks Agree? Fixing Benchmark Evaluation with BenchBench," exposes the flaws in current BAT practices. The researchers analyzed 50 popular LLM benchmarks and found that arbitrary choices in how benchmarks are compared can drastically skew the results. Picking different comparison benchmarks, using too few models for testing, or even setting arbitrary thresholds for "agreement" can lead to vastly different conclusions. This makes it difficult to trust any single benchmark and even harder to compare LLMs fairly. To fix this mess, the researchers propose a set of best practices for BAT, including using an "aggregate reference benchmark" (combining multiple benchmarks for a more stable comparison), a data-driven threshold for agreement, more models for testing, and reporting results at different levels of detail. They've even created a Python package called BenchBench that automates these best practices. BenchBench not only streamlines benchmark comparisons but also hosts a dynamic leaderboard (BenchBench-Leaderboard) that ranks benchmarks based on their agreement with chosen references. This creates a more transparent and reliable system for evaluating LLMs and allows researchers to make informed choices. The implications of this research are far-reaching. By establishing a fairer and more consistent evaluation process, BenchBench can foster trust in LLM benchmarks and accelerate the development of truly robust and intelligent AI systems. The quest for better AI evaluation isn't over, but BenchBench is a significant step towards a more level playing field. Future work should focus on tackling reliability concerns within individual benchmarks and exploring how best to use these improved BAT methods to identify obsolete benchmarks that are no longer relevant to the evolving capabilities of LLMs.