FedSPZO splits an LLM into two blocks and uses more perturbations on the smaller second block, reusing activations to cut zero-order federated finetuning compute by up to 3x versus prior zero-order methods.
SL o RA : Federated parameter efficient fine-tuning of language models
1 Pith paper cite this work. Polarity classification is still indexing.
1
Pith paper citing it
fields
cs.LG 1years
2025 1verdicts
CONDITIONAL 1representative citing papers
citing papers explorer
-
Efficient Zeroth-Order Federated Finetuning of Language Models on Resource-Constrained Devices
FedSPZO splits an LLM into two blocks and uses more perturbations on the smaller second block, reusing activations to cut zero-order federated finetuning compute by up to 3x versus prior zero-order methods.