RL over Commodity Networks: Overcoming the Bandwidth Barrier with Lossless Sparse Deltas

Researchers have developed SparrowRL, a new system that makes reinforcement learning (RL) training for large language models practical over standard internet connections. It achieves this by sending only the ~1% of model parameters that actually change each step, instead of the full model. This reduces transfer payload by 79x for an 8B model and improves throughput by up to 9.5x over standard methods on wide-area networks, narrowing the performance gap with expensive dedicated clusters to under 9%.