Skip to content
conftrace_
2026 EMNLP EMNLP 2026

Unveiling Latent Cognitive Bias Constructs in Large Language Models

AbstractAs Large Language Models (LLMs) are increasingly adopted in decision-making, there is a growing need to better understand the behavioural patterns underlying their reasoning. Evaluation is one of the primary approaches used to study model behaviour. However, most existing LLM evaluations rely on task-based benchmarks that summarise performance through aggregate scores. While effective for comparing model performance, such approaches provide limited insight into the latent behavioural tendencies that drive biased reasoning. In this paper, we propose a construct-based evaluation framework for studying cognitive bias in LLMs grounded in psychometric evaluation. We systematically perturb prompts across controlled conditions to analyse reasoning behaviour. From model responses, we derive behavioural features capturing susceptibility, robustness, correction, and variability, and apply latent variable modelling to infer underlying behavioural constructs. Our findings show that cognitive bias in LLMs is structured and dynamic, varying across bias type, model families, and scale.

More from EMNLP 2026

All EMNLP 2026 papers →