Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Deconstructing Self-Bias in LLM-generated Translation Benchmarks

Wenda Xu, Sweta Agrawal, Vilém Zouhar, Markus Freitag, Daniel Deutsch

TL;DR

This work formalizes and quantifies self-bias in the LLM-as-a-benchmark paradigm for translation, showing that both testset generation and evaluation contribute to bias and that bias is amplified in into-English directions for low-resource languages. It identifies two bias mechanisms—dialect-induced and translatability-induced—and demonstrates that limited source-text diversity underpins much of the effect. Through ablations and diversity-focused analyses, the paper shows that increasing source-text diversity can mitigate self-bias, though translation asymmetry persists due to resource and lexical-diversity disparities. Despite the bias challenges, LLM-as-a-benchmark can still reliably rank open-source models and remain useful for rapid iteration in certain contexts, provided biases are accounted for or mitigated. Overall, the work highlights practical implications for the design of scalable benchmarks and suggests avenues for reducing bias via data diversity and controlled evaluation protocols.

Abstract

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.

Deconstructing Self-Bias in LLM-generated Translation Benchmarks

TL;DR

This work formalizes and quantifies self-bias in the LLM-as-a-benchmark paradigm for translation, showing that both testset generation and evaluation contribute to bias and that bias is amplified in into-English directions for low-resource languages. It identifies two bias mechanisms—dialect-induced and translatability-induced—and demonstrates that limited source-text diversity underpins much of the effect. Through ablations and diversity-focused analyses, the paper shows that increasing source-text diversity can mitigate self-bias, though translation asymmetry persists due to resource and lexical-diversity disparities. Despite the bias challenges, LLM-as-a-benchmark can still reliably rank open-source models and remain useful for rapid iteration in certain contexts, provided biases are accounted for or mitigated. Overall, the work highlights practical implications for the design of scalable benchmarks and suggests avenues for reducing bias via data diversity and controlled evaluation protocols.

Abstract

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.

Paper Structure

This paper contains 23 sections, 5 equations, 1 figure, 13 tables.

Table of Contents

  1. Introduction
  2. Preliminaries and Self-Bias Definition
  3. Experimental Setup
  4. Does the use of LLM-as-a-benchmark introduce self-bias?
  5. Where does the self-bias stem from?
  6. How does source text generation impact self-bias?
  7. Source text generation leads to self-bias
  8. What makes source text generation exhibit bias?
  9. Underlying Causes: Limited Diversity and Quality in Synthetic Low-Resource Language Source Texts
  10. Improving on diversity can reduce Self-bias
  11. Impact of Translation Direction on Self-Bias
  12. Why does translation asymmetry exist for self-bias?
  13. Merits of LLM-as-a-benchmark
  14. LLM-as-a-benchmark can still benefit open source models
  15. Related work
  16. ...and 8 more sections

Figures (1)

  • Figure 1: The left panel illustrates the Type-Token Ratio (TTR) of source texts (three LLMs vs. FLORES human references), indicating higher lexical diversity in human texts and distinct TTR profiles among LLMs. Bottom right, pairwise Cohen's D values exceeding 0.5 quantify these TTR distributional differences as moderate to substantial. Top right, estimated degeneration ratios across two language directions reveal differential degeneration levels in the three models during source text generation (Degeneration is defined as 4-gram repeating more than 10 times).