IC-899GPT-4, when used as a blind pairwise evaluator, exhibits the same style-over-factuality preference as human crowdworkers

Arnav Gudibande, Eric Wallace, Charlie Victor Snell, Xinyang Geng, Hao Liu, Pieter Abbeel, Sergey Levine, Dawn Song

SourceThe False Promise of Imitating Proprietary Language Models

The paper uses GPT-4 to conduct blind pairwise comparisons between ChatGPT outputs and the authors' imitation model outputs, following the same procedure as their Mechanical Turk evaluations. GPT-4's ratings show the same trends as human raters: it rates imitation models as competitive with ChatGPT, particularly when the imitation models produce stylistically similar (confident, well-structured) outputs. The authors note this implies that LLMs may replicate human-like cognitive biases in evaluation, and that GPT-4 may be a viable cheap proxy for human evaluation on some tasks, albeit with a slightly larger absolute preference for ChatGPT's outputs.

Evidence

correlational

Caveat

The observation is brief and secondary to the paper's main focus on imitation models. No detailed analysis of GPT-4's evaluation mechanism is provided; the claim rests on the qualitative observation that GPT-4's rating trends match those of crowdworkers.

Model

GPT-4 / ChatGPT4 / GPT-4 Code Interpreter / GPT-4 Technical Report

Extraction

automatic-extraction