Simran Arora

Simran Arora researches AI systems and efficient model execution. Her work connects GPU programming with the question of whether language models can generate kernels that are both correct and fast. She is a co-author of ParallelKernelBench, an evaluation suite and execution harness for multi-GPU kernel generation.

In work associated with Together AI, she studies the communication and scheduling problems that appear when computation spans GPUs. Her analysis separates producing a correct program from improving its performance, examining where additional samples and agent feedback help and where those gains plateau.