A Comprehensive Evaluation of Cognitive Biases in LLMs

TL;DR

This work tackles the problem of cognitive biases in large language models by proposing a scalable, general-purpose test framework and a large- scale benchmark. It introduces a four-entity, three-function framework to generate, decide, and estimate bias-related test cases, coupled with 200 managerial scenarios to yield 30,000 tests across 30 biases in 20 LLMs. The study demonstrates that cognitive biases are pervasive across models, with biases present in all models for most biases and complex patterns of bias linked to model size and architecture, while providing a public dataset and framework to drive mitigation and future research. The results underscore the need for careful LLM selection in high-stakes decision-making and offer a practical, extensible methodology for bias evaluation across domains beyond managerial contexts.

Abstract

We present a large-scale evaluation of 30 cognitive biases in 20 state-of-the-art large language models (LLMs) under various decision-making scenarios. Our contributions include a novel general-purpose test framework for reliable and large-scale generation of tests for LLMs, a benchmark dataset with 30,000 tests for detecting cognitive biases in LLMs, and a comprehensive assessment of the biases found in the 20 evaluated LLMs. Our work confirms and broadens previous findings suggesting the presence of cognitive biases in LLMs by reporting evidence of all 30 tested biases in at least some of the 20 LLMs. We publish our framework code to encourage future research on biases in LLMs: https://github.com/simonmalberg/cognitive-biases-in-llms

Figures (13)

• Figure 1: An LLM changes its answer as the framing of the decision changes, indicating the susceptibility of the LLM to the Framing Effect.
• Figure 2: Our overall test pipeline comprises four steps: for each test case, it (1) takes a scenario and a test case with two templates as input, (2) samples two instances of the templates by inserting suitable values into all template gaps, (3) lets a decision LLM choose one option for each template instance, and (4) uses the corresponding metric to estimate the final bias value.
• Figure 3: Cumulative distribution of cosine similarity scores for the datasets.
• Figure 4: Metric codomain for scale $\sigma_1 = \{1, 2, ..., 7\}$, $y_1 = y_2 = 0$ and different values of parameter $k$.
• Figure 5: The plot shows the absolute biasedness (i.e., the strength of the biasedness, independent of direction) of models in relation to their size (bubble diameter) and Chatbot Arena score (as a measure of general capability). When no such score is available, we take the mean of the other models' scores and mark the model with a '*'.