Reproduce “Towards Understanding Bias in Synthetic Data for Evaluation” 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
Hossein A. Rahmani, Varsha Ramineni, Emine Yilmaz, Nick Craswell, Bhaskar Mitra
An investigation into LLM-generated synthetic test collections for Information Retrieval evaluation reveals inherent biases: LLMs are more lenient judges, inflating absolute system performance metrics by assigning higher relevance scores, and showing a preference for systems aligned with their own generative patterns. While absolute scores are affected, the relative ranking of systems often remains moderately to highly correlated with human-based evaluations.
Key insights
- LLM-generated queries are typically longer and exhibit different vocabulary patterns compared to human-generated queries.
- LLM judges are generally more lenient than human judges, assigning a higher proportion of 'related' or higher relevance scores.
- Synthetic test collections tend to overestimate the absolute performance of IR systems and can introduce a bias favoring LLM-based retrieval systems.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.