Story perspectives

MIT Reveals AI Models Solve Problems Like Humans

11/21/2025

46 8

Summary

Story summary

Researchers at MIT's McGovern Institute for Brain Research report reasoning-focused large language models process information in a human-like way.
They note that humans and models take longer on difficult problems, like the Abstraction and Reasoning Challenge (ARC), which involves inferring transformations.
Engineers train these models with reinforcement learning, rewarding correct answers and guiding stepwise problem solving.
This approach improves accuracy and yields more consistent results, even if it raises runtime.

1 of 1

Story summary

Researchers at MIT's McGovern Institute for Brain Research report reasoning-focused large language models process information in a human-like way.
They note that humans and models take longer on difficult problems, like the Abstraction and Reasoning Challenge (ARC), which involves inferring transformations.
Engineers train these models with reinforcement learning, rewarding correct answers and guiding stepwise problem solving.
This approach improves accuracy and yields more consistent results, even if it raises runtime.