SourcePhysics of Language Models: Part 3.2, Knowledge Manipulation
The paper tests GPT-4 and GPT-3.5 on a task where a 4-character Chinese idiom has one character masked and the model must fill it in. The accuracy depends dramatically on which position is masked: for GPT-4, accuracy ranges from 17.6% (position 1) to 90.6% (position 4); for GPT-3.5, from 9.4% to 56.7%. The same three visible characters serve as evidence in all four cases, but the model's ability to recover the missing character depends on where the gap sits in the sequence. This demonstrates that the model's response to the same evidence depends on its position in the input.