Scientists uncover problems in artificial intelligence “reasoning”

Large language models (LLMs) to be precise. A new preprint published on the arXiv website sheds light on the limitations of large language models (LLMs) in analogue reasoning tasks. Research shows that large language models, such as GPT models, perform suboptimally compared to humans, especially when solving letter string analogy problems with standard alphabets. DiscussScientists have uncovered problems in artificial intelligence 'reasoning'© Ferra

Moreover, when presented with counterfactual alphabets, LLMs show a decrease in accuracy, exhibiting different error patterns from human ones, indicating a lack of abstract thinking required for the advanced AI that many companies are now striving for.

The comparison between human intelligence and LLM, although insightful, faces difficulties. LLMs work in the digital realm, which limits the possibility of direct comparison with digitally expressed human abilities. Although LLMs demonstrate mastery of tasks, they lack emotional attachment, remorse, and awareness of consequences—characteristics inherent in human cognition.

Despite their mastery of digital tasks, LLMs pose unique challenges in the field of digital security. Their capabilities extend to creating fake content, raising concerns about the spread of misinformation and cybersecurity threats. As LLMs navigate complex digital environments, ensuring they are used responsibly is paramount to reducing the risks of digital contamination and malicious activity.

To address these challenges, efforts to align LLM capabilities with human cognition and development are critically needed. measures to protect against digital risks. Understanding the parallels between LLM qualifications and human cognitive performance can inform the development of AI security and digital memory management strategies that protect against the spread of malicious digital content and cyber threats, experts say.