The Vulnerability of Language Model Benchmarks: Do They Accurately Reflect True LLM Performance?
Sourav Banerjee, Ayushi Agarwal, Eishkaran Singh
TL;DR
This paper interrogates the reliability of large language model benchmarks by dissecting how evaluation pipelines are vulnerable to hacking, contamination, and bias. It surveys historical and contemporary evaluation regimes, highlighting how overfitting to benchmarks, data leakage, adversarial perturbations, and flawed human/LLM judging distort apparent progress. The authors argue for dynamic, domain-aware evaluation frameworks that resist manipulation and track provenance, proposing governance structures and zero-day, zero-shot protocols to better approximate true language understanding and real-world performance. The work underscores the practical impact: model developers may chase scores rather than genuine competence, so robust, adaptive benchmarks are essential for trustworthy advancement.
Abstract
The pursuit of leaderboard rankings in Large Language Models (LLMs) has created a fundamental paradox: models excel at standardized tests while failing to demonstrate genuine language understanding and adaptability. Our systematic analysis of NLP evaluation frameworks reveals pervasive vulnerabilities across the evaluation spectrum, from basic metrics to complex benchmarks like GLUE and MMLU. These vulnerabilities manifest through benchmark exploitation, dataset contamination, and evaluation bias, creating a false perception of progress in language understanding capabilities. Through extensive review of contemporary evaluation approaches, we identify significant limitations in static benchmark designs, human evaluation protocols, and LLM-as-judge frameworks, all of which compromise the reliability of current performance assessments. As LLM capabilities evolve and existing benchmarks become redundant, we lay the groundwork for new evaluation methods that resist manipulation, minimize data contamination, and assess domain-specific tasks. This requires frameworks that are adapted dynamically, addressing current limitations and providing a more accurate reflection of LLM performance.