Tag
This paper identifies weighting noise in LLM judges for multi-stakeholder tasks and proposes DecompR, a method that decouples utility estimation from aggregation using counterfactually calibrated weights.
A multi-institution survey proposes a three-layer trust framework to align technical, clinical, and human-centered requirements for trustworthy AI in mental-health support.
OpenAI and UC Berkeley's workshop on Confidence-Building Measures for Artificial Intelligence brought together stakeholders to develop strategies for mitigating geopolitical risks from foundation models, identifying six key CBMs including crisis hotlines, incident sharing, model transparency, content provenance, red teaming, and dataset sharing.