When More Attention Makes Motor Diagnosis Worse
A motor fault diagnosis study shows why complementary attention mechanisms can outperform isolated modules and larger networks, with limited additional computation.
A motor fault diagnosis study shows why complementary attention mechanisms can outperform isolated modules and larger networks, with limited additional computation.
A cyber-social severity framework captures who and what may be affected by industrial attacks, but higher scores alone do not establish better risk prioritization.
A traffic-routing benchmark shows why more reliable prediction bounds can produce worse route choices, and how operators should evaluate the difference.
Why retrieving relevant conversation history can undermine instruction adherence, and what role-aware memory changes for enterprise AI assistants.
A new attribution method reveals how benchmark shortcuts and extra inference computation can inflate AI agent improvements—and how to evaluate what genuinely transfers.
AgenticRV shows why preventing unsafe AI-agent actions depends on enforceable tool boundaries, sufficient evidence, and task-specific authorization—not simply adding monitors.
CALLIOPE shows how LLMs can build interpretable, differentially private topic hierarchies without embedding-based clustering, at a substantial inference cost.
VisualEvolve shows how version selection and structured failure evidence can improve visual AI agents without retraining their underlying models.
Why audio language models can retain vocal evidence yet follow contradictory text, and what targeted interventions reveal about the failure.
ABot-ER shows how reliable evaluation, supervision repair, and reusable training data can matter more than another round of model optimization.