The Comparative Trap: Pairwise Comparisons Amplifies Biased Preferences of LLM Evaluators
This addresses bias in LLM-based evaluation for natural language generation, which is crucial for fair assessments in AI research, though it is incremental as it builds on existing evaluation methods.
The paper tackles the problem of biased preferences in LLM evaluators, showing that pairwise evaluation amplifies biases like favoring verbosity, while pointwise evaluation is less susceptible. It introduces PRePair, a method that reduces bias and improves performance on benchmarks such as LLMBar and MT-Bench.
As large language models (LLMs) are increasingly used as evaluators for natural language generation tasks, ensuring unbiased assessments is essential. However, LLM evaluators often display biased preferences, such as favoring verbosity and authoritative tones. Our empirical analysis reveals that these biases are exacerbated in pairwise evaluation, where LLMs directly compare two outputs and easily prioritize superficial attributes. In contrast, pointwise evaluation, which assesses outputs independently, is less susceptible to such bias because each output is judged in isolation. To address the limitations of the pairwise evaluation, we introduce a novel evaluation method, PRePair, which integrates pointwise reasoning within a pairwise framework. PRePair effectively alleviates biased preference, improving performance on the adversarial benchmark (LLMBar) while outperforming pointwise evaluation on the standard benchmark (MT-Bench).