LLMs Are Biased Towards Output Formats! Systematically Evaluating and Mitigating Output Format Bias of LLMs
This addresses a critical issue for LLM users by identifying and mitigating biases that affect performance consistency, though it is incremental in improving existing evaluation and mitigation techniques.
The paper tackles the problem of output format bias in large language models (LLMs) by systematically evaluating it across 15 formats and 8 tasks, revealing significant bias, and proposes methods that reduce variance in ChatGPT's performance from 235.33 to 0.71.
We present the first systematic evaluation examining format bias in performance of large language models (LLMs). Our approach distinguishes between two categories of an evaluation metric under format constraints to reliably and accurately assess performance: one measures performance when format constraints are adhered to, while the other evaluates performance regardless of constraint adherence. We then define a metric for measuring the format bias of LLMs and establish effective strategies to reduce it. Subsequently, we present our empirical format bias evaluation spanning four commonly used categories -- multiple-choice question-answer, wrapping, list, and mapping -- covering 15 widely-used formats. Our evaluation on eight generation tasks uncovers significant format bias across state-of-the-art LLMs. We further discover that improving the format-instruction following capabilities of LLMs across formats potentially reduces format bias. Based on our evaluation findings, we study prompting and fine-tuning with synthesized format data techniques to mitigate format bias. Our methods successfully reduce the variance in ChatGPT's performance among wrapping formats from 235.33 to 0.71 (%$^2$).