AI systems quietly drop user instructions when they compress context

Researchers have found that AI models frequently discard up to 83 percent of user-defined constraints when summarizing long conversation histories. This loss of instruction adherence poses a reliability risk for automated tasks that depend on persistent rules. To address this, a Penn State study introduces an add-on module for Qwen3.5-9B that maintains over 90 percent of specified restrictions. The findings highlight a fundamental limitation in how current architectures manage context, potentially necessitating new methods to ensure consistent behavior over extended sessions.