AI Model Improves Prediction Using Self-Generated Data
Researchers from Tel Aviv University and Stanford University show a computer model can sharpen its predictions by training on data it creates itself.
Updated on Nov. 4, 2026 in Artificial Intelligence
Live Poll
Do you think AI will make your daily life better over the next few years?
A study published on the scientific website arXiv describes two computer models trained together, with one generating symbol sequences as learning exercises and the other learning to predict the next symbol. The trained model was tested on language, images, sounds, music, computer code, and DNA sequences.
Why it matters
The approach suggests AI systems could learn more efficiently when they encounter real-world data, reducing reliance on externally supplied training material. Its limits are notable, since the method cannot teach facts about the real world without exposure to real-world information.
The study reports that prediction ability rose as more computing power was applied during training, but no benchmark scores or exact performance gains are given. The stub does not specify how the self-generated data compared in quality with externally supplied training data.
The players
Tel Aviv University
A major public research university in Tel Aviv, Israel, which issued the statement on the study.
Stanford University
A private research university in California whose researchers co-led the work with Tel Aviv University.
The details
The first model gradually improved at creating learning exercises that helped the second model make better predictions. Because the method relies only on self-generated data, it cannot teach facts about the real world without real-world exposure.
Timeline
Nov. 4, 2026: Tel Aviv University issued a statement on the AI self-learning study.
The Tech Race
The work extends the self-play tradition that drove earlier game-playing AI systems into prediction across language, images, sounds, music, code, and DNA. It positions self-generated training as a possible route to more data-efficient learning as competition over AI training data intensifies.
If the approach matures, AI tools could become more capable at adapting to new data without retraining on large external datasets, which may affect the speed and cost of future AI products. The stub gives no timeline for commercial availability, so users should not expect immediate changes.
The takeaway
Self-generated training is a promising direction for making AI learn more efficiently, but its inability to supply real-world facts means it will complement rather than replace exposure to real-world information. Readers should watch for independent validation before assuming broad gains.
Further reading
Explore more coverage in Artificial Intelligence, where related research on machine learning methods is tracked.
Live Poll
Do you think AI will make your daily life better over the next few years?
