Reproduce “Beyond the Surface: Measuring Self-Preference in LLM Judgments” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Zhi-Yuan Chen, Hao Wang, Xinyu Zhang, Enrui Hu, Yankai Lin
Researchers at Chen et al. developed the DBG score, a new metric to accurately measure self-preference bias in LLM judges by disentangling it from response quality. Their findings indicate that larger models exhibit less bias, and that shared training data and response style normalization can help mitigate this bias across various model types and datasets.
Key insights
- Both pre-trained and instruction-tuned models exhibit self-preference bias, with instruction-tuned models not necessarily showing more bias.
- Larger models consistently demonstrate less self-preference bias than smaller models.
- Response style normalization and shared training data can help mitigate self-preference bias, while attention analysis suggests judges naturally allocate more attention to their own responses.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.