Reproduce “Word Overuse and Alignment in Large Language Models: The Influence of Learning from Human Feedback” 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
Tom S. Juzek, Zina B. Ward
Research from Florida State University establishes a causal link between Learning from Human Feedback (LHF) and the overuse of specific lexical items in large language models. The study developed an automated procedure to identify words favored by LHF-tuned models and experimentally validated that human evaluators subtly prefer text variants containing these words.
Key insights
- Learning from Human Feedback (LHF) is a direct causal factor for the lexical overuse observed in LLMs, driven by the subtle preferences of human evaluators.
- The process of "alignment" via LHF can inadvertently lead to a form of "misalignment," where the model reflects the specific lexical preferences of LHF workers rather than broader user expectations.
- The linguistic preferences of LHF workers, often from the Global South, significantly influence the global language output of AI, raising sociolinguistic and ethical questions about language standardization.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.