Instructed to Bias: Instruction-Tuned Language Models Exhibit Emergent Cognitive Bias

Itay Itzhak, Gabriel Stanovsky, Nir Rosenfeld, Yonatan Belinkov

Research output: Contribution to journalArticlepeer-review


Recent studies show that instruction tuning (IT) and reinforcement learning from human feedback (RLHF) improve the abilities of large language models (LMs) dramatically. While these tuning methods can help align models with human objectives and generate highquality text, not much is known about their potential adverse effects. In this work, we investigate the effect of IT and RLHF on decision making and reasoning in LMs, focusing on three cognitive biases—the decoy effect, the certainty effect, and the belief bias—all of which are known to influence human decisionmaking and reasoning. Our findings highlight the presence of these biases in various models from the GPT-3, Mistral, and T5 families. Notably, we find a stronger presence of biases in models that have undergone instruction tuning, such as Flan-T5, Mistral-Instruct, GPT3.5, and GPT4. Our work constitutes a step toward comprehending cognitive biases in instruction-tuned LMs, which is crucial for the development of more reliable and unbiased language models.1.

Original languageEnglish
Pages (from-to)771-785
Number of pages15
JournalTransactions of the Association for Computational Linguistics
StatePublished - 4 Jun 2024
Externally publishedYes

Bibliographical note

Publisher Copyright:
© 2024 Association for Computational Linguistics.


Dive into the research topics of 'Instructed to Bias: Instruction-Tuned Language Models Exhibit Emergent Cognitive Bias'. Together they form a unique fingerprint.

Cite this