Fed-FSTQ: Fisher-Guided Token Quantization for Efficient Federated LLM Fine-Tuning
A new method, Fed-FSTQ, addresses communication bottlenecks in federated fine-tuning of large language models (LLMs) on edge devices. The system uses a lightweight Fisher proxy to estimate token sensitivity, enabling importance-aware token selection and non-uniform mixed-precision quantization. This approach allocates higher fidelity to informative signals while suppressing redundant transmission, reducing per-round payloads under heterogeneous bandwidth and intermittent participation. Fed-FSTQ is model-agnostic and designed for communication-efficient federated learning in non-IID data regimes.
Key facts
- Fed-FSTQ is a Fisher-guided token quantization system for federated LLM fine-tuning.
- It uses a lightweight Fisher proxy to estimate token sensitivity.
- It couples importance-aware token selection with non-uniform mixed-precision quantization.
- The method reduces per-round payloads in non-IID regimes.
- It is model-agnostic.
- It targets communication-efficient federated learning on edge devices.
- The work is published on arXiv with ID 2604.25421.
- The approach addresses straggler-limited uplink communication under heterogeneous bandwidth.
Entities
Institutions
- arXiv