A team of researchers has presented a new language model that achieves 94% accuracy on complex mathematical problems, beating GPT-5 in this area for the first time. The model, developed by a European academic consortium, uses a hybrid reasoning architecture that combines transformer networks with symbolic modules designed specifically for formal reasoning.
The results, published in the journal Nature Machine Intelligence, have surprised the scientific community with the efficiency of the approach: the model requires only 30% of GPT-5’s parameters to obtain better results on logical-mathematical reasoning tasks. The researchers have announced that the model’s code and weights will be published in open access in the coming weeks, which could significantly accelerate research in this field.
