LLM Benchmarks Are Junk Science

Author(s): Kaushik Rajan Originally published on Towards AI. An Oxford review of 445 benchmarks found 84% lack basic statistical testing. Models score 90% on standard tests but 2% on unseen problems. A 5-question smell test for any benchmark claim. Over the past year, I’ve evaluated more than sixty artificial intelligence (AI) tools for production use. Every vendor came armed with benchmark numbers: leaderboard rankings, accuracy percentages, scores on tests with impressive names. The numbers looked scientific. Decimal points. Bold-faced winners. Tables sorted by fractions of a percent. Image by the author.This article critiques the reliability of benchmarks used in assessing artificial intelligence tools, revealing that a substantial majority fail to employ statistical testing, leading to misleading claims about model performance. It discusses the disconnect between benchmark results and real-world performance, highlighting issues such as contamination of test data and the use of vague definitions in benchmarks. The author emphasizes the need for more rigorous evaluation standards and proposes a checklist to assess benchmark validity. Read the full blog for free on Medium. Join thousands of data leaders on the AI newsletter. Join over 80,000 subscribers and keep up to date with the latest developments in AI. From research to projects and ideas. If you are building an AI startup, an AI-related product, or a service, we invite you to consider becoming a sponsor. Published via Towards AI