IC-681GPT-3.5 exhibits positional bias when judging which of two LLM responses is superior

Yidong Wang, Zhuohao Yu, Wenjin Yao, Zhengran Zeng, Linyi Yang, Cunxiang Wang, Hao Chen, Chaoya Jiang, Rui Xie, Jindong Wang, Xing Xie, Wei Ye, Shikun Zhang, Yue Zhang

SourcePandaLM: An Automatic Evaluation Benchmark for LLM Instruction Tuning Optimization

The paper reports an observed inherent bias in GPT-3.5 where its evaluation of which response is better depends on the order in which the two responses are presented in the prompt. Even with carefully designed prompts, GPT-3.5's judgment conflicts when the response order is swapped. The authors mitigate this by filtering out inconsistent training samples and by swapping the order at inference time, revising conflicting results to 'tie'. No specific magnitude or proportion of affected samples is reported.

Evidence
observational
Caveat
The paper does not quantify the proportion of samples affected by the bias or the magnitude of the order effect; it is described qualitatively as 'inherent' and handled via filtering.
Model
GPT-3.5 / ChatGPT-3.5
Concepts
Positional bias
Related findings
IC-682, IC-683
Extraction
automatic-extraction