Our paper on speeding up Log-Sum-Exp has been accepted at NeurIPS 2026! We introduce faster CPU and GPU kernels for this key operation underlying softmax, with potential benefits for AI models that rely heavily on it.

• CPU: Approximate integer arithmetic accelerates the computation, tested on particle filtering.
• GPU: Kernel fusion reduces memory traffic and launch overhead, tested on RLHF log-probability computation and probabilistic circuit inference.