Dissecting Multiplication in Transformers: Insights into LLMs

Transformers struggle with basic arithmetic despite advanced language abilities: Large language models like GPT-4 perform poorly on seemingly simple tasks like integer multiplication, raising concerns about their reliability for applications requiring mathematical reasoning and their safe deployment in real-world scenarios.

The root cause is difficulty managing intermediate steps: The research identifies that transformers fail at multiplication primarily because they struggle to track "carryovers" (the numbers carried over between digit calculations) and store intermediate results—fundamental requirements for multi-step arithmetic that humans handle automatically.

Models use an inefficient parallel processing strategy: Rather than solving multiplication the way humans do sequentially, transformers attempt to decompose the task into multiple parallel subtasks, which creates bottlenecks and errors when results need to be combined.

Targeted improvements can dramatically boost performance: By addressing these specific limitations, researchers achieved over 99.9% accuracy on 5-digit multiplication with a small transformer model, even outperforming GPT-4—demonstrating that understanding why AI fails is key to fixing it.

Better understanding of AI limitations improves both safety and capability: This work contributes to responsible AI development by making transformer reasoning more interpretable and showing that targeted architectural improvements can address fundamental weaknesses, supporting more trustworthy AI systems.