🏢 University of Southern California

Multicalibration post-processing isn’t always necessary; models often implicitly achieve it, especially calibrated ones. For uncalibrated models, though, it significantly improves fairness.

Transformers surprisingly learn second-order optimization methods for in-context linear regression, achieving exponentially faster convergence than gradient descent!

PGLDM, a novel algorithm, accurately identifies shared and private dynamics in two generalized-linear time series, improving model accuracy and enabling lower-dimensional latent state representations.

This paper proves the minimax optimal U-calibration error is Θ(√KT) for online multiclass prediction, resolving an open problem and showing logarithmic error is achievable for specific loss functions.

DIVA: Evolutionary task generation for robust, adaptable AI agents in complex simulators.

A new dynamic subgroup identification strategy, using covariate-adjusted response-adaptive randomization, efficiently identifies best-performing subgroups in clinical trials, improving resource alloca…

DOPPLER, a novel low-pass filter, significantly enhances differentially private (DP) optimizer performance by reducing the impact of privacy noise, bridging the gap between DP and non-DP training.