Today's trail

Pathology foundation models (PFMs) support a wide range of histopathology tasks on whole-slide images through multiple instance learning (MIL). Despite recent progress, these models remain sensitive to cross-hospital distribution shifts: performance often degrades when a task…

Camera design heavily influences perception task performance, yet cameras are typically designed by domain experts through costly experimentation, in isolation from the tasks they serve. Existing joint-design methods either rely on differentiable simulation limited to…

Diffusion models generate samples through a multi-step denoising trajectory, yet condition every step on the same instruction, so the trajectory cannot be addressed at the level of individual operations. We introduce Dancing Diffusion, which assigns each span of timesteps a…

VLMs for medical image interpretation are typically evaluated on clean inputs, despite the progressive evidence loss that can occur in high stakes clinical settings. Existing robustness studies show that performance can degrade while confidence remains high, but do not…