Reproduce “Mitigating Label Biases for In-context Learning” 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
Yu Fei, Yifan Hou, Zeming Chen, Antoine Bosselut
Researchers from UC Irvine, EPFL, and ETH Zurich conceptualized and empirically demonstrated "domain-label bias," a novel type of bias in large language models where task-specific vocabulary influences label predictions in in-context learning. Their proposed Domain-context Calibration method, which mitigates vanilla-label, context-label, and especially domain-label biases, improved Macro-F1 scores by up to 37% on challenging datasets, preventing models from performing at random chance.
Key insights
- Large Language Models exhibit a "domain-label bias," where associations between task-specific vocabulary and labels learned during pre-training can systematically bias predictions.
- This domain-label bias can cause LLMs to perform no better than random on certain tasks, significantly hindering in-context learning capabilities.
- Calibrating with random words from the *in-domain* task corpus, rather than generic content-free tokens or random English words, is crucial for effectively mitigating domain-label bias.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.