Chinese researchers at Microsoft have developed a new technique to improve the ability of large language models to solve mathematical problems. The technique of learning from mistakes is similar to that of humans. It is no coincidence that the model is called “Learning from Mistakes” (LeMa).

Researchers from Microsoft Research in Beijing and Xi’an Jiaotong University unveiled their self-described groundbreaking strategy on Cornell University’s website. According to their work, the language model is trained to correct its own errors, resulting in improvement in reasoning skills. Interestingly, Google also introduced some innovations last week to help students with more difficult math problems.

Human learning processes

The researchers were inspired by human learning processes in which students learn from mistakes to improve their final performance. “A student who was unable to solve a math problem will learn from the mistakes he or she made. In this way, the student learns to correct the mistake,” say the authors of the work. They applied the same principle to language models (LLMs), using “error correction data pairs” generated by GPT-4 to refine them.

Specifically, the researchers had models create “broken” reasoning paths for mathematical word problems. GPT-4 identified these errors, explained them, and corrected the reasoning. The researchers then used this adjusted data to further train the original models. “LeMa continuously improves performance and is an important step towards AI systems that can learn from their mistakes and improve,” the research team said.

Open source approach

The broader AI community can review the researchers’ work and start doing it themselves. All data, including code and models, are available via GitHub. This step should enable further advances in machine learning. Developing a model that applies reason like humans could revolutionize sectors such as healthcare, finance, but also autonomous vehicles, where error correction and continuous learning are of course crucial.

According to Venture Beat, there is exciting potential in the field of artificial intelligence. As machines become more adept at learning from their mistakes, we are moving toward a future where AI can outperform humans in complex situations, it is said. We’ll leave it to the editors to decide whether we should all be so happy about it.