Reproduce “Causal-Guided Active Learning for Debiasing Large Language Models” 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

Li Du, Zhouhao Sun, Xiao Ding, Yixuan Ma, Yang Zhao, Kaitao Qiu, Ting Liu, Bing Qin

Researchers from Harbin Institute of Technology and BAAI developed Causal-Guided Active Learning (CAL), a framework that automatically identifies and mitigates dataset biases in large language models using causal invariance and in-context learning. CAL improves LLM generalizability and reduces harmful content by discovering explainable and previously unreported bias patterns.

Key insights

  • Causal invariance can effectively distinguish true semantic relationships from spurious, correlative biases in large language models.
  • Large language models (LLMs) can autonomously identify and summarize their own underlying bias patterns, including those previously unknown.
  • In-context learning (ICL) provides an effective and efficient method for debiasing LLMs, avoiding the computational cost and generalizability issues associated with fine-tuning.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.