Language models show human-like content effects on reasoning

By Ishita Dasgupta, Andrew Kyle Lampinen, Stephanie C. Y. Chan, Antonia Creswell, Dharshan Kumaran, James L. McClelland, Felix Hill. 2022

230 citations recorded by the source corpus.

Reasoning is a key ability for an intelligent system. Large language models (LMs) achieve above-chance performance on abstract reasoning tasks, but exhibit many imperfections. However, human abstract reasoning is also imperfect. For example, human reasoning is affected by our real-world knowledge and beliefs, and shows notable"content effects"; humans reason more reliably when the semantic content of a problem supports the correct logical inferences. These content-entangled reasoning patterns play a central role in debates about the fundamental nature of human intelligence. Here, we investigate whether language models $\unicode{x2014}$ whose prior expectations capture some aspects of human knowledge $\unicode{x2014}$ similarly mix content into their answers to logical problems. We explored this question across three logical reasoning tasks: natural language inference, judging the logical validity of syllogisms, and the Wason selection task. We evaluate state of the art large language models, as well as humans, and find that the language models reflect many of the same patterns observed in humans across these tasks $\unicode{x2014}$ like humans, models answer more accurately when the semantic content of a task supports the logical inferences. These parallels are reflected both in answer patterns, and in lower-level features like the relationship between model answer distributions and human response times. Our findings have implications for understanding both these cognitive effects in humans, and the factors that contribute to language model performance.

Subfields: Artificial Intelligence, Computer Vision and Pattern Recognition, Large Language Models, Natural Language Processing

How researchers see this paper

The researchers whose trusted networks rank it highest, strongest advocate first. Every researcher gets a different order.

In Artificial Intelligence

In Computer Vision and Pattern Recognition

In Large Language Models

In Natural Language Processing

Citations in the corpus

Augmented Language Models: a Survey (2023, background): “Moreover, we can expect this phenomenon to be amplified as LMs reason in quite a similar manner to humans (Dasgupta et al., 2022), making it even harder to detect mistakes.”
Faithful Reasoning Using Large Language Models (2022, background): “Presented with a context of relevant knowledge and a question, an LM may base its answer on information encoded in its weights rather than prioritising the information present in the context (Dasgupta et al., 2022).”
Using Large Language Models to Simulate Multiple Humans and Replicate Human Subject Studies (2022, background): “Dasgupta et al. (2022) test LMs on abstract reasoning problems and ﬁnd that “such models often fail in situations where humans fail – when stimuli become too abstract or conﬂict with prior understanding of the world.””
Mind's Eye: Grounded Language Model Reasoning through Simulation (2022, background): “, 2022) is able to elicit LM reasoning in a more explicit way; however, its performance is reported to be sensitive to the quality of human annotated reasoning steps (Lampinen et al., 2022; Dasgupta et al., 2022) (i.”
Large Language Models Are Reasoning Teachers (2022, unknown): “We can hypothesize that factors like content bias may play a role here; language models have been found to fail depending on context and content of the task, in a way similar to human reasoners (Dasgupta et al., 2022).”
PersonaLLM: Investigating the Ability of Large Language Models to Express Personality Traits (2023, background): “…that, as the size of LLMs increases, LLMs demonstrate emerging abilities as agents (Andreas, 2022) and exhibit human-like behavior in reasoning (Dasgupta et al., 2022; Webb et al., 2023; Binz and Schulz, 2023; Aher et al., 2023; Wong et al., 2023), role-playing (Wang et al., 2023b; Shao et…”
Looped Transformers as Programmable Computers (2023, background): “LLMs can also perform algorithmic tasks and reasoning through in-context learning (ICL), as shown in several works, such as (Nye et al., 2021; Wei et al., 2022c; Lewkowycz et al., 2022; Wei et al., 2022b; Dasgupta et al., 2022; Chung et al., 2022).”
Large language models in healthcare and medical domain: A review (2023, background): “This advancement could significantly improve applications in genomics and personalized medicine, enabling these models to predict individual responses to treatments based on genetic profiles, thereby advancing the precision of medical interventions.”
Towards Reasoning in Large Language Models: A Survey (2022, background): “Finally, we refer the reader to Dohan et al. (2022)’s position paper on language model cascade , which presents a unifying framework for understanding chain-of-thought prompting and research in this line.”
Playing repeated games with large language models (2023, unknown): “Previous research efforts have explored various approaches to analyze LLMs, such as employing cognitive psychology tools 51,64 and even adopting a computational psychiatry perspective 52 .”
Defending ChatGPT against jailbreak attack via self-reminders (2023, unknown): “The collection and filtering process for the further 37 jailbreak prompts in the independent training set is detailed in Supplementary Information section 1.2.”
Language Models Are Greedy Reasoners: A Systematic Formal Analysis of Chain-of-Thought (2022, result): “As in our approach, Dasgupta et al. (2022) speciﬁcally looked at whether LLMs can reason in ﬁctional or counterfactual settings and found that reasoning ability is indeed negatively affected in these settings.”
Reasoning or Reciting? Exploring the Capabilities and Limitations of Language Models Through Counterfactual Tasks (2023, background): “Nevertheless, LMs struggle with reasoning with premises that are inconsistent with common sense (Dasgupta et al., 2022; Yu et al., 2023; Tang et al., 2023).”
Using cognitive psychology to understand GPT-3 (2022, methodology): “” Large language models, such as GPT3, have previously been studied using vignette-like problems (10, 15, 16).”
The rise and potential of large language model based agents: a survey (2025, unknown): “They conducted psychology experiments about judgment and decision-making to test agent systems [413, 414, 416, 417].”
MARS: Multi-hop Adaptive Retrieval and SPARQL Generation for KGQA (2026, unknown): “Although LLMs show promising results in reasoning [51], integrating structured data into KGQA remains challenging [61].”