More about HKUST
The Cost of Collaboration: A Survey of Efficiency-Aware Multi-LLM Systems
PhD Qualifying Examination
Title: "The Cost of Collaboration: A Survey of Efficiency-Aware Multi-LLM
Systems"
by
Mr. Haochen SHI
Abstract:
Multi-LLM systems extend single-agent LLMs by composing multiple model calls,
agents, roles, memories, tools, and runtime resources into a coordinated
computation. This horizontal scaling path can improve task performance through
decomposition, specialization, parallel search, context isolation, and
verification, but it also introduces substantial resource overhead in tokens,
latency, tool calls, memory, model loading, serving throughput, and training
or rollout cost. This survey studies how recent multi-LLM systems improve
efficacy while controlling these costs.
We organize the literature into a two-level taxonomy. At the harness level,
we review methods that optimize the logical structure of the system:
capability allocation selects which model, role, expert, or subteam should
execute each unit of work; work-graph control shapes task decomposition,
branching, scheduling, pruning, and stopping; information-state control
decides what each model call or agent sees, retrieves, stores, and shares;
and external-action control covers emerging methods for budgeting tools,
APIs, code execution, environments, and human clarification. At the device
level, we review runtime methods that make these harness decisions efficient
on concrete hardware, including workflow-aware scheduling, multi-LLM and
multi-LoRA serving, KV-cache and context-state optimization, placement,
batching, and training-time rollout efficiency.
Across these domains, we emphasize quality-resource trade-offs, baseline
selection, and end-to-end accounting. The survey provides a map for
designing multi-LLM systems that are not only more capable, but also
cost-aware, deployable, and empirically accountable.
Date: Monday, 22 June 2026
Time: 10:00am - 12:00pm
Venue: Room 3494
Lift 25/26
Committee Members: Dr. Yangqiu Song (Supervisor)
Prof. Nevin Zhang (Chairperson)
Dr. May Fung