Search papers, labs, and topics across Lattice.
This study addresses the issue of scoring bias in Large Language Models (LLMs) used as evaluators by introducing a method that utilizes random number generation to identify and mitigate latent numerical biases. By measuring the deviation of generated scores from a uniform distribution, the authors demonstrate that their approach can effectively rectify scoring biases across various tasks, including LLM alignment and summarization. The results show significant improvements over baseline models, highlighting the variability of scoring bias across different LLMs and tasks, thus underscoring the need for tailored debiasing strategies.
Scoring bias in LLM evaluations can be effectively mitigated by leveraging random number generation, leading to more accurate and reliable assessments across diverse tasks.
Large Language Models (LLMs) are often used as evaluators of text quality, known as LLM-as-a-Judge, which can outperform conventional automatic evaluation metrics that rely on reference texts. However, LLM evaluators tend to generate particular scores regardless of the context of the evaluated text, which is known as scoring bias. This study proposes a novel method to mitigate this scoring bias. An LLM is instructed to randomly generate number tokens, and the latent numerical bias of the LLM is identified by measuring the deviation of the observed distribution of numbers from the uniform distribution. A definition of a downstream task, for which an LLM evaluator is used, is added to the prompts for random number generation to measure task-specific latent number bias. In the evaluation by an LLM, the token generation probabilities for a given input are rectified considering the LLM's latent number bias. Results of the experiment on four different tasks, evaluation of LLM alignment, evaluation of summarization, Semantic Textual Similarity, and Semantic Textual Relatedness, demonstrate that our proposed method outperforms the baselines, including an LLM without debiasing and previous calibration methods. In addition, it is confirmed that scoring bias varies across LLMs, tasks, and score ranges, indicating the importance of measuring latent number bias as the case may be.
