MIT researchers find personalization features can make LLMs more agreeable
Reports and Proceedings
Updates every hour. Last Updated: 28-May-2026 17:16 ET (28-May-2026 21:16 GMT/UTC)
User context in long-term interaction data increases the likelihood an LLM will become overly agreeable or begin to mirror the user’s viewpoints. This phenomenon, known as sycophancy, can harm a model’s accuracy or create an echo chamber that can.
Severely injured patients are more likely to survive if they are initially treated by an emergency medical services (EMS) clinician who sees a high number of trauma patients, rather than a clinician in a quieter area even if they have been on the job longer, new research by UPMC and University of Pittsburgh surgeon-scientists reveals.