A Systematic Comparison of Syllogistic Reasoning in Humans and Language Models

TL;DR

This work investigates whether PaLM 2 and Llama 2 language models perform deductive syllogistic reasoning in a way that aligns with normative logic or mirrors human biases. Using a controlled dataset of $64$ syllogism types evaluated with zero-shot chain-of-thought prompting, the authors compare LM responses to human data and map LM behavior into the Mental Models Theory framework via mReasoner, revealing that larger models often surpass humans in accuracy yet retain systematic, human-like biases such as figural effects and syllogistic fallacies. The study finds that LM reasoning becomes more deliberative with scale, yet deliberation only partly accounts for accuracy, indicating a dissociation between correctness and the underlying reasoning process. These results have implications for AI interpretability and cognitive modeling, showing that LMs can achieve high logical performance while still inheriting and sometimes overcoming human biases present in their training data, with practical impact on the design and evaluation of logic-enabled AI systems.

Abstract

A central component of rational behavior is logical inference: the process of determining which conclusions follow from a set of premises. Psychologists have documented several ways in which humans' inferences deviate from the rules of logic. Do language models, which are trained on text generated by humans, replicate such human biases, or are they able to overcome them? Focusing on the case of syllogisms -- inferences from two simple premises -- we show that, within the PaLM2 family of transformer language models, larger models are more logical than smaller ones, and also more logical than humans. At the same time, even the largest models make systematic errors, some of which mirror human reasoning biases: they show sensitivity to the (irrelevant) ordering of the variables in the syllogism, and draw confident but incorrect inferences from particular syllogisms (syllogistic fallacies). Overall, we find that language models often mimic the human biases included in their training data, but are able to overcome them in some cases.

Figures (21)

• Figure 1: The zero-shot chain-of-thought prompt we use to assess LM syllogistic reasoning. The different parts of the prompt are grouped together for illustration purposes only; see also Figure \ref{['fig:prompts']} in the Appendix for a purely textual representation of the prompt.
• Figure 2: Accuracy of PaLM 2 models, humans (red), and random guessing (grey). Random guessing accuracy differs by syllogism as some syllogisms have more than one valid conclusion. Syllogisms are partitioned into variable ordering (by row) and ordered by decreasing human accuracy from left to right. The top right inset shows the average accuracy across all syllogisms. Syllogisms are identified with the letters of the moods of the premises (Table \ref{['table:moods-variables']}, left) and the number associated with their variable ordering (Table \ref{['table:moods-variables']}, right).
• Figure 3: Correlation between PaLM 2 models' predictions and human predictions. The oracle here is a logically correct reasoner that samples a response at random from all valid responses; the correlation of such an oracle with humans is relatively low as it does not mimic human errors.
• Figure 4: Variable ordering effects in PaLM 2 models and humans. Left: The marginal probabilities of A-C and C-A ordered conclusions. Right: The magnitude of the variable ordering effect (the absolute value of the difference between the C-A probability and the A-C probability).
• Figure 5: Right: Each syllogism plotted by accuracy (y-axis) and entropy (x-axis) and the regression line relating the two. Dashed lines black lines show the residuals for each of the top three human syllogistic fallacies. Left: The result of correlating PaLM 2's regression residuals with residuals estimated from human data.