Breaking the Benchmark: Revealing LLM Bias via Minimal Contextual Augmentation
✖
Investigated whether large language models truly reduce social bias by applying minimal, meaning-preserving augmentations to the BBQ benchmark, extending evaluation to underexplored categories like ageism and disability, revealing latent bias patterns and model-specific robustness differences.
- Authors: Kaveh Eskandari, Mahammed Kamruzzaman, Me, Gene Louis Kim, Vasanth Sarathy and Ninareh Mehrabi (Supervisor)
- Paper (arXiv)
RADAR: Risk-Aware Dynamic Adaptive Returns
✖
RADAR learns per-stock regularized Sharpe q-scores, balancing exploration via trade counts using UCB-RSSR; ensures anytime average-cost compliance for observation-agnostic, capital-efficient portfolio adaptation, risk-managed, profitable growth.