Researchers discover a shortcoming that makes LLMs less reliable
Reports and Proceedings
Updates every hour. Last Updated: 19-Apr-2026 10:16 ET (19-Apr-2026 14:16 GMT/UTC)
MIT researchers find that large language models sometimes mistakenly link certain grammatical sequences to specific topics, and then rely on these learned patterns when answering queries. This phenomenon can cause LLMs to fail unexpectedly on new tasks and could be exploited by adversarial agents to trick an LLM into generating harmful content.
Researchers from University of Shanghai for Science and Technology, China have developed a twisted double-layer graphene plasmonic metasurface that achieves unprecedented confinement of terahertz waves into nanoscale volumes, theoretically enabling fingerprint detection of molecular monolayers as thin as 1 nm. This system overcomes the critical challenge in terahertz sensing where the long wavelength (hundreds of micrometers) weakly interacts with nanoscale molecules. By engineering acoustic plasmon nanocavities through precise twist angles between graphene layers, the team demonstrated a mode volume as small as 10⁻¹³λ₀³ and sensitivity 48 times higher than conventional single-layer graphene and non-twist double-layer graphene structures. The platform provides a new insight for ultra-strong light-matter interaction at terahertz frequencies and opens possibilities for single-molecule spectroscopy and on-chip biosensing applications.
While most organizations address cybersecurity issues with technology and surveillance, a new study from the University of Vaasa, Finland, argues that empathy may be a more effective defence. Emmanuel Anti's doctoral research explores insider deviance, and how understanding the human elements related to it can lead to stronger, more sustainable cybersecurity practices. His dissertation proposes an empathetic security model grounded in design thinking, encouraging organisations to co-create cybersecurity policies with employees, focusing on understanding their needs, motivations and emotional well-being.
Researchers have discovered that ferroelectric fluids can harness an overlooked transverse electrostatic force (TEF) to rise over 80 mm, without magnets or high voltages. By exploiting the fluid’s spontaneous polarization and exceptionally high dielectric constant, they achieved a strong TEF, previously thought unattainable in conventional electrostatics. This breakthrough enables creation of a lightweight, magnet-free motor, opening possibilities for compact, energy-efficient actuators and suggesting a transformative approach to converting electrical energy into mechanical motion at low voltages.
A European consortium of researchers has developed a new decision making framework for health professionals which they hope will transform the care of patients with chronic obstructive pulmonary disease (COPD).A European consortium of researchers has developed a new decision making framework for health professionals which they hope will transform the care of patients with chronic obstructive pulmonary disease (COPD).
While artificial intelligence (AI) is designed to enhance the quality of our lives, there is an increase in AI-mediated harm. Since AI systems operate semi-autonomously, neither their developers nor users typically intend or foresee such harm. Thus, it seems absurd to hold AI systems and stakeholders responsible, giving rise to a responsibility gap. To address this gap, researchers have now developed new notions of responsibility tailored to cover unintended and unforeseen consequences caused by AI.