Fully local synthesis on rural edge hardware

Determine whether a compact on-device language model running on the same edge hardware can match the cross-modal reasoning quality of a cloud large language model over structured clinical evidence and establish the latency it can achieve on CPU-only rural hardware, thereby eliminating the cloud round trip under total uplink failure.

Background

The proposed cloud–edge system performs modality-specific perception locally but delegates cross-modal reasoning and clinical-summary generation to a cloud LLM. Although the system transmits only approximately 6.5 KB of structured evidence per case, synthesis still depends on network connectivity.

The paper identifies fully local synthesis with a small LLM as a natural extension. Such a model could generate summaries during complete uplink failure, but its ability to preserve the reasoning quality of a cloud model—and its practical CPU-only latency on rural edge hardware—has not been established. The problem is therefore concerned with both quality parity and deployment-time feasibility.

References

Whether a compact on-device model can match the cross-modal reasoning quality of a cloud LLM over structured evidence---and at what latency on CPU-only rural hardware---is an open question we leave to future work.

— A Cloud-Edge System for Multimodal Clinical Screening in Resource-Constrained Rural Settings  (2608.12745 - Ting et al., 13 Aug 2026) in Appendix, Section "Privacy and Fully-Local Synthesis," paragraph "Toward Fully Local Synthesis"