The paper tests whether released language models can perform simple logical operations on stored factual knowledge. GPT-4 is asked to classify whether a celebrity was born in an even month (binary) and to compare two people's birth dates. Without explicitly generating the intermediate value (the month or the year) first, GPT-4 performs at or near chance level. The same failure is confirmed for GPT-4o and Llama-3.1-405B as of October 1, 2024. The paper's controlled synthetic experiments show this is not a training-data issue: even with 25,000 QA samples (far exceeding the theoretical minimum) the model cannot learn the classification without CoT.
Evidence
correlational
Key metric
GPT-4: 50.7% correct on 'born in an even month?' (binary); 52.3% on comparing birth dates among individuals born 1900-1910; 71.1% for 1900-1950; 81.6% for all pairs. GPT-4o and Llama-3.1-405B: 'still fail' (Figure 5, no specific numbers in text).
Caveat
The paper notes these are the simplest single-step manipulations; more complex composed functions would be even harder. The GPT-4o and Llama-3.1-405B results are qualitative ('still fail') without specific accuracy numbers in the text.