AI that reasons with itself learns better, faster, and with less data

Talking to yourself is often seen as a very human habit. We do it when we’re stressed, planning something complex, or trying to make sense of a tough decision.

That quiet inner voice helps us organize thoughts, weigh options, and move forward. Now, researchers are discovering that this same idea might unlock a new level of intelligence in machines.

A new study published in Neural Computation by scientists at the Okinawa Institute of Science and Technology (OIST) suggests that artificial intelligence learns faster and becomes more flexible when it’s trained to “talk to itself.”

Instead of relying only on memory or raw data processing, these AI systems use a form of internal dialogue—described by researchers as subtle “mumbling”—combined with short-term working memory. The result? Smarter learning across a wide range of tasks.

According to the research team, learning isn’t just about how an AI system is built. It’s also about how that system interacts with itself during training. By structuring training data in a way that encourages self-interaction, the AI begins to develop better learning strategies. This mirrors how humans reflect internally before acting, especially in unfamiliar situations.

When researchers tested this approach, the benefits were clear. AI models that combined inner speech with working memory adapted more easily to new problems, handled multitasking better, and performed well even when data was limited.

Tasks that required holding multiple pieces of information—such as reversing sequences or reconstructing patterns—showed especially strong improvements.

One of the most exciting implications is generalization. Humans are great at applying what they know to new situations. AI, on the other hand, often struggles outside of narrow training scenarios. By using inner speech and structured working memory, these systems rely more on general rules than memorized examples, bringing them closer to human-like learning.

Even more promising, this method works with sparse data. In a world where training massive models requires enormous datasets and energy, a lightweight alternative could be a game changer.

Next, the researchers plan to test these ideas in messy, real-world environments—where noise, unpredictability, and constant change are the norm.

Beyond advancing AI, this work may also deepen our understanding of human learning itself, and help build robots capable of functioning in homes, farms, and other dynamic spaces we live in every day.