research21h
Quantifying and Auditing LLM Evaluation via Positive--Unlabeled Learning
Researchers frame LLM evaluation under selective human supervision as a positive-unlabeled learning problem. They propose a method to quantify and audit biases in LLM-as-a-Judge systems, finding systematic issues like verbosity bias. The approach helps builders assess LLM reliability in real-world scenarios. This work informs strategies to improve LLM evaluation.
Key takeaways
- LLM-as-a-Judge systems show systematic biases like verbosity bias.
- Positive-unlabeled learning can quantify LLM evaluation biases.
- Method helps assess LLM reliability in real-world scenarios.