HybridFlow: Adaptive Task Scheduling for Fast and Token-Efficient LLM Inference in Edge-Cloud Collaboration
Published in ICML, 2026
HybridFlow introduces an adaptive task scheduling framework for efficient LLM inference across edge-cloud environments, optimizing both latency and token usage.
Framework
Figure 1: HybridFlow operates in two tightly coupled stages: (1) Task Decomposition & Parallel Execution — an edge-side planner decomposes a complex query into a DAG of interdependent subtasks; (2) Resource-Aware Subtask Routing — a lightweight learned router adaptively assigns each subtask to either the edge SLM or the cloud LLM based on a benefit–cost utility score.
Recommended citation: Dong, Jiangwen, Jiayu Li, and Wanyu Lin. "HybridFlow: Adaptive Task Scheduling for Fast and Token-Efficient LLM Inference in Edge-Cloud Collaboration." ICML 2026.
Download Paper