RBCorr: Response Bias Correction in Language Models

Researchers unveiled RBCorr, a low-cost method to correct response bias in language models. Tested on 12 open-weight models, it effectively eliminated option preference biases in yes-no, entailment, and multiple-choice questions. The correction boosted model performance, showing pre-correction bias is prevalent. The technique's success depends heavily on the specific model, dataset, and prompt format used. It helps ensure benchmark results better reflect a model's true capabilities.