Proceedings of the Annual Meeting of the Cognitive Science Society

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.