Researchers have developed SplitLite, a method that reduces the high communication costs associated with split learning in on-device large language models. This is done by exploiting the low effective rank structure of consecutive-epoch activation and gradient residuals. The method involves transmitting only quantized truncated singular value decomposition residual factors, which significantly reduces both activation uplink and gradient downlink traffic. Experiments show that SplitLite can reduce communication costs by up to 93.5% without performance degradation. --- Why it matters: This matters because it addresses a significant limitation of split learning in on-device large language models: the high communication costs associated with exchanging activations and gradients between clients and servers. By reducing these costs, SplitLite enables more efficient fine-tuning of LLMs on-device, which is crucial for applications where data is sensitive or cannot be easily transferred. Source: https://arxiv.org/abs/2608.23018