Tag
GradRepair-ODE introduces a reliability framework for certifying and repairing gradients in Neural ODE training to address numerical stability issues in scientific machine learning and generative models.
This paper audits NAVSIM v2.2's defensive driving scoring, showing that a shared reference-conditioned forgiveness rule combined with numerical instability can cause actor-blind probes to outrank human replay, undermining the benchmark's validity.
This article demonstrates that using stochastic rounding for BF16 optimizer state can match FP32 performance because unbiased errors cancel over time, whereas round-to-nearest stalls due to compounding bias. An experiment with an MLP shows BF16+SR achieves similar loss to FP32 while using less memory.