一作SIGIR 2026Poster· Information Retrieval (Short)
The LLM Effect on IR Benchmarks: A Meta-Analysis of Effectiveness, Baselines, and Contamination
Moritz Staudinger · Wojciech Kusa · Allan Hanbury
摘要
Benchmark collections have long enabled controlled comparison and cumulative progress in Information Retrieval (IR). However, prior meta-analyses show that reported effectiveness gains often fail to accumulate, in part due to weak or outdated baselines. Large language models (LLMs) are increasing...