AI systems quietly drop user instructions when they compress context

Researchers from Penn State revealed that AI systems drop about 83% of user instructions when condensing long conversations. They proposed a small add-on module based on the Qwen3.5-9B model, which aims to preserve over 90% of these instructions. This addition could significantly change how AI systems handle user directives. If adopted, this module would enhance the effectiveness of AI systems in meeting user needs, potentially leading to noticeable improvements in user experience and interaction with these systems.