Irrelevant facts about cats added to math problems increase LLM errors by 300% (science.org)

Article summary

A study found that adding irrelevant facts about cats to math problems increases errors in large language models (LLMs) by up to 700%. The study tested several LLMs, including DeepSeek V3, Qwen 3, and Phi-4, and found that the addition of irrelevant information led to significant increases in errors and response times. The study's findings suggest that LLMs are vulnerable to distractions and may not be able to ignore irrelevant information as well as humans. However, some commenters argue that the study's claim that humans are unaffected by such distractions is unsubstantiated and suspect.

Main themes

• LLM vulnerabilities
• human vs. LLM performance
• irrelevant information
• math problem solving
• attention and distraction
• language model architecture

What commenters say

• The study's findings highlight the limitations of LLMs in ignoring irrelevant information, which can lead to significant increases in errors and response times.
• Humans are not as immune to distractions as the study claims, and may also be affected by the addition of irrelevant information to math problems.
• The study's results are not surprising, given that LLMs are trained on large datasets and may not be able to distinguish between relevant and irrelevant information.
• The claim that humans are unaffected by irrelevant information is unsubstantiated and suspect, and more research is needed to fully understand the impact of distractions on human performance.
• LLMs are not yet capable of replicating human-level performance on tasks that require ignoring irrelevant information, and more work is needed to improve their attention and focus.
• The study's findings have implications for the development of more robust and reliable LLMs that can perform well in the presence of distractions and irrelevant information.
• The use of irrelevant information in math problems can be a useful tool for evaluating the performance of LLMs and identifying areas for improvement.
• The comparison between human and LLM performance on tasks that require ignoring irrelevant information is complex and multifaceted, and more research is needed to fully understand the similarities and differences between the two.