Large language models (LLMs) are impressive at many tasks, but math word problems remain a challenge. They often stumble because they rely on superficial keyword matching rather than genuine mathematical reasoning. Think about how you solve a math problem—you probably draw upon similar problems you've encountered in the past and adapt the solution. New research explores this very idea by investigating how LLMs can learn by analogy, using computational graphs. A computational graph essentially maps out the underlying mathematical relationships within a problem. By retrieving examples with similar computational graphs, LLMs can be prompted with problems that follow the same reasoning path. This allows the LLM to ‘see’ the correct logic and apply it to the new problem. The researchers found this method significantly improved LLM performance on six different math word problem datasets, achieving up to a 6.7% boost in accuracy. This improvement was especially prominent in smaller LLMs, suggesting this method could be crucial for making more compact AI models more powerful.Interestingly, the research showed that the LLM didn’t need a huge amount of training data to learn by analogy effectively. Even with a smaller training dataset, performance saw a noticeable jump. This points to the potential of creating highly effective LLMs without needing massive amounts of annotated data. The researchers also explored a clever workaround for the problem of manually creating computational graphs, which can be time-consuming and expensive. They used other LLMs like GPT-4 to generate similar problems, effectively automating the creation of training data. While this automated approach wasn't quite as effective as using human-labeled data, it still outperformed standard LLMs. This work opens exciting avenues for future research. By allowing LLMs to learn by analogy, we can equip them with the tools to solve complex reasoning tasks, not just in math but potentially in any field that requires structured thinking.