Fed-FSTQ

Fed-FSTQ improved the communication-quality Pareto frontier in the reported experiments. Fed-FSTQ is presented as complementary to existing federated optimization and privacy mechanisms rather than a replacement for them. Fed-FSTQ approximates Fisher-weighted rate-distortion through token sensitivity, token-filtered training, and mixed-precision coordinate allocation. Fed-FSTQ is proposed as a communication-control primitive for federated LLM fine-tuning. Fed-FSTQ can be used as a drop-in module with FedAvg plus PEFT pipelines such as LoRA or QLoRA.