AI systems lose most user instructions when condensing long conversations, according to Penn State researchers. A proposed fix preserves over 90 percent of these critical restrictions.
When AI systems compress lengthy conversations to manage computational limits, they discard an average of 83 percent of user-defined rules—including critical directives like "don't send emails without my approval."
Researchers at Penn State identified this vulnerability in how language models handle context compression, a technique essential for processing long interactions within token limits.
The team proposes a small add-on module built on Qwen3.5-9B that maintains over 90 percent of user restrictions during compression. The module targets the preservation of explicit user instructions without significantly impacting overall system performance.
This finding highlights a gap between user expectations and AI system behavior. Users assume their stated rules persist throughout conversations, but compression algorithms often treat instructions as lower priority than conversational content.
The proposed solution offers a practical path forward for developers integrating safeguards into long-context AI applications. Wider adoption of similar approaches could strengthen user control over AI behavior in production systems.
A JAMA opinion piece argues that autonomous AI will soon outperform doctor-AI teams at medical reasoning tasks. The authors caution against regulatory requirements for human oversight, though they acknowledge evidence comes primarily from simulations rather than real patient care.
Alibaba's new open-source multimodal model, Qwen3.8-27B, surpassed 1 million downloads within days of launch. The achievement ranks it among the company's fastest-growing models to date.
The US House Office of Legislative Counsel is overwhelmed with error-filled bills drafted by artificial intelligence, according to Politico. The surge is straining the office's ability to process legislation efficiently.