SplitQ-PEFT: Quantum-Enhanced PEFT for Communication-Efficient Split Learning

초록

Split learning offers a practical framework for privacy-aware distributed training, but adapting large language models in such settings still requires substantial communication and optimization cost. We propose SplitQ-PEFT, a quantumenhanced PEFT framework for communication-efficient split learning. By introducing compact quantum-smashed representations, SplitQ-PEFT significantly reduces the per-step communication overhead while retaining competitive adaptation performance. Across multiple benchmarks, SplitQ-PEFT requires the fewest trainable parameters and achieves the lowest communication overhead among the compared PEFT baselines, while maintaining competitive performance on generative tasks. Further entropy and visualization analyses show that the quantumsmashed representation remains compact yet expressive. Our results suggest that SplitQ-PEFT is an effective approach for communication-efficient and parameter-efficient LLM adaptation in split learning environments.

키워드