Reproduce “B-score: Detecting biases in large language models using response history” 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
An Vo, Mohammad Reza Taesiri, Daeyoung Kim, Anh Totti Nguyen
Researchers introduce the B-score, an unsupervised metric quantifying LLM bias by comparing single-turn and multi-turn conversational response patterns. They find that many LLM biases, such as preferential digit selection, are correctable when models can self-reflect on their output history, leading to more uniform distributions and improved answer verification accuracy by up to 9.3%.
Key insights
- Many observed LLM biases are artifacts of single-turn prompting and are correctable through multi-turn self-reflection, where models observe their own response history.
- LLMs possess an inherent capacity for balanced output generation when given appropriate context about their own response patterns.
- The B-score effectively differentiates between correctable biases (e.g., in random number generation) and more persistent, genuine preferences.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.