Deconstructing Self-Bias in LLM-generated Translation Benchmarks
This exposes a fundamental problem for researchers and practitioners relying on automated benchmarks to evaluate LLMs, revealing that current methods may produce misleading rankings.
The paper identifies a critical flaw in LLM-generated translation benchmarks where models systematically favor themselves, particularly in low-resource language-to-English translation tasks, showing that bias originates from both test data generation and evaluation methods and is amplified when combined.
As large language models (LLMs) begin to saturate existing benchmarks, automated benchmark creation using LLMs (LLM as a benchmark) has emerged as a scalable alternative to slow and costly human curation. While these generated test sets have to potential to cheaply rank models, we demonstrate a critical flaw. LLM generated benchmarks systematically favor the model that created the benchmark, they exhibit self bias on low resource languages to English translation tasks. We show three key findings on automatic benchmarking of LLMs for translation: First, this bias originates from two sources: the generated test data (LLM as a testset) and the evaluation method (LLM as an evaluator), with their combination amplifying the effect. Second, self bias in LLM as a benchmark is heavily influenced by the model's generation capabilities in the source language. For instance, we observe more pronounced bias in into English translation, where the model's generation system is developed, than in out of English translation tasks. Third, we observe that low diversity in source text is one attribution to self bias. Our results suggest that improving the diversity of these generated source texts can mitigate some of the observed self bias.