Topic

LLM robustness

Study of how far model performance holds up when inputs are noisy, adversarial or unlike the material a model was benchmarked on.

Current clusters

build1 publisher

VeriSim's injected patient noise costs seven open-weight models 15 to 25 accuracy points

VeriSim injects recall gaps, low health literacy and stigma-driven non-disclosure into simulated patients, and its authors report seven open-weight models shedding 15 to 25 diagnostic accuracy points while conversations run 34 to 55 percent longer.

Publishers:arxiv.org

Reality

Evidence32
Adoption: Insufficient
Hype gap+28
Incentives62
Confidence44