Sangmin Lee

Generative MARL teachers fail by interpolating coordination modes; a novel optimal transport approach forces them to commit to distinct strategies, boosting student performance.

Consumer GPUs can now achieve up to 1.66x faster MoE inference by bypassing NCCL's CPU bottlenecks with a novel DMA-driven communication design.

Speculative decoding can be significantly accelerated by training models to anticipate verification outcomes, leading to faster and more efficient inference.