UniK preprint reports stronger retrieval results with a 70-billion-parameter model

Researchers report that UniK, a system that organizes and retrieves information from different kinds of data, helped an open-source 70-billion-parameter model achieve 76% accuracy on a government-data question-answering test, compared with 47% for GPT-5. They also report 77.9% accuracy on a medical question-answering test without task-specific fine-tuning. UniK enriches source material, builds multiple search indexes and combines their results before the model answers. The authors propose using the same approach for data used to train physical AI systems, but the preprint reports its measured results in digital AI domains rather than a demonstrated robotics deployment.