Artificial intelligence systems learn new tasks more reliably when they generate their own internal self-talk while problem-solving.
The addition of a private language layer reshapes how the system transfers knowledge from one problem to the next, allowing the system to adapt without retraining.
Self-talk improves learning
The multitasking learning system was most powerful when internal self-talk worked in parallel with short-term memory when actively solving problems.
Jeffrey Quicer of Okinawa Institute of Science and Technology Graduate University (OIST) has demonstrated that a model that verbalizes goals can maintain internal direction even when switching between tasks.
The system maintained performance on new tasks by applying previous operations rather than relearning them.
There were clear limits to performance gains, raising the question of whether internal self-talk could keep up with increasing task complexity.
Role of memory slot
Humans rely on working memory to keep instructions valid when deciding what to do next.
In the OIST simulations, a model with two memory slots kept the task goals separate from the pattern details during execution.
This separation allowed the system to reverse the sequence and regenerate it without mixing steps from competing tasks.
The single-slot design had a tendency to override itself, placing an upper limit on handling unfamiliar combinations.
Training using self-talk
In the new simulation, the system practiced a memory task while generating short, autonomous utterances during operation.
These utterances served as inner speech, a type of silent monologue that indicated goals and helped the network pause before choosing its next step.
“By structuring the training data in a way that teaches the system to talk to itself, we showed that learning is shaped not only by the architecture of the AI system, but also by the interaction dynamics built into the training procedure,” said Kweisser.
Repeated self-talk turns out to be most useful when a task involves many steps, allowing you to maintain planning stability while multitasking.
Learning with AI systems
Many AI systems only improve after consuming large datasets, forcing researchers to explore more lightweight learning methods.
Self-talk gave the model additional practice signals, and memory slots made the goals of each task available over long sequences.
This combination supported generalization, using past learning to deal with new situations, even if the training set remained sparse.
Since the tests were conducted within a controlled task, this approach still needs to be demonstrated in a noisy environment like a real robot.
Stable goals, changing patterns
The system is designed to learn and switch tasks, not to predict words or carry on a conversation.
Instead of ingesting huge text datasets, the model was trained based on structured tasks that require stable goals amidst changing patterns.
Rather than encode meaning, its internal language served as a control signal, organizing steps and priorities.
This contrast shows how flexibility in learning derives from structure and memory, not just scale.
Layers handle different speeds
Within the network, separate layers began to handle different time scales, giving the system a built-in sense of order.
The fast layer tracks the changing pattern stream, while the slow layer keeps the task goal stable from beginning to end.
This timing division allowed the memory module to store content without the burden of control decisions.
When its internal rhythm was not formed, the same architecture lost its edges and functioned like a mere photocopy machine.
The system predicts and adjusts
The task was performed under active inference, a control method that updates predictions to reduce errors at each step.
Each step started by predicting what would happen next, and the system adjusted if the results did not match.
The self-talk output is fed back into the system as a new input, allowing the controller to recheck its goals during execution.
Because the framework prioritizes stable predictions, unexpected events can force rapid updates and tax the same memory resources.
Humans also talk to themselves
A 2015 review tracked inner speech from childhood to adulthood and linked it to planning and self-control.
People often use that silent talk to keep their goals in mind and correct themselves when they get distracted.
Inner speech also supports emotional regulation and intentional decision-making in everyday life, as it allows us to label our emotions and choices.
The AI model borrowed that idea in a narrow form, so readers should not treat it as spirit.
Can robots use self-talk?
Rigorous training quickly fails as domestic and agricultural machinery faces changing instructions, slippery objects, and frequent interruptions.
“By exploring phenomena like inner language and understanding the mechanisms of such processes, we can gain fundamental new insights into human biology and behavior,” Kweisser concluded.
If robots learn this way, fewer script rehearsals may be needed, but engineers will need to carefully test for safety.
By combining self-talk and working memory, mock learners were able to control not only the final answer, but also their own steps.
Whether that advantage holds up in noisy real-world settings, and even on physical machines, remains an open question.
The research will be published in a journal neural computing.
—–
Like what you read? Subscribe to our newsletter for fascinating articles, exclusive content and the latest updates.
Check us out on EarthSnap, the free app from Eric Ralls and Earth.com.
—–
