{"pk":49671,"title":"Estimating and Correcting Yes-No Bias in Language Models","subtitle":null,"abstract":"When presented with a yes-no question, humans tend to say 'yes' regardless of the ground truth. This 'yes-bias' can be attributed either to the social pressure to agree with an interlocutor or simply to the tendency to mimic the distribution of the input data. Here, we estimate 'yes-no' response bias in language models (LMs), with the goal of distinguishing the two theories, and explore two strategies for bias correction. We develop two yes-no question datasets derived from existing world knowledge datasets, and test 16 open-weight LMs. We find that LMs often show response bias on yes-no questions, but that it is highly variable, deviating from bias observed in humans. We further present a novel bias correction method, which eliminates bias and improves model performance. Evidence of non-humanlike response bias in LMs informs us on the source of yes-bias in humans, and the efficacy of our bias correction method holds promise for LM evaluation.","language":"eng","license":{"name":"","short_name":"","text":null,"url":""},"keywords":[{"word":"Artificial Intelligence; Language and thought; Machine learning; Reasoning; Computational Modeling"}],"section":"Papers with Poster Presentation","is_remote":true,"remote_url":"https://escholarship.org/uc/item/2c04k26b","frozenauthors":[{"first_name":"Om","middle_name":"","last_name":"Bhatt","name_suffix":"","institution":"Georgia Institute of Technology","department":""},{"first_name":"Anna","middle_name":"","last_name":"Ivanova","name_suffix":"","institution":"Georgia Institute of Technology","department":""}],"date_submitted":null,"date_accepted":null,"date_published":"2025-01-01T13:00:00-05:00","render_galley":null,"galleys":[{"label":"PDF","type":"pdf","path":"https://journalpub.escholarship.org/cognitivesciencesociety/article/49671/galley/37633/download/"}]}